Detecting Sewer Infiltration: Rainfall-Flow Lag Analysis and Multi-Method I&I Detection for Municipal Wastewater Systems

Abstract

Inflow and infiltration wastes energy. Every litre of stormwater or groundwater that enters a sanitary sewer must be pumped, aerated, and treated at full cost — consuming electricity that produces no public health benefit. This post presents a four-method detection toolkit for quantifying I&I across a multi-plant municipal wastewater system: cross-correlation lag analysis to identify response timing, hydrograph decomposition to separate dry weather flow from rainfall-derived contributions, conductivity-based dilution tracing to independently confirm extraneous flow fractions, and machine learning to predict RDII from weather inputs including degree-day snowmelt. The methodology requires only SCADA influent flow data, freely available Environment Canada rainfall and temperature records, and monthly energy bills — no asset management data, no pipe condition surveys, no GIS. The output is a plant-level I&I severity ranking with energy cost quantification and remediation prioritization that directs limited rehabilitation budgets where they will deliver the highest return.

1. Introduction — The Energy Cost Nobody Tracks

Wastewater utilities know they have inflow and infiltration. Every operator has watched influent flow spike after a summer thunderstorm or climb steadily through a spring thaw. The wet weather surge is visible on every SCADA screen, documented in every annual report, and discussed at every operations meeting. What is rarely quantified is the energy penalty — the actual cost, in kilowatt-hours and dollars, of pumping, aerating, and treating water that should never have entered the sanitary sewer system.

This is not a small number. Wastewater treatment consumes 1,000 to 3,000 kWh per million gallons, depending on facility size, process configuration, and effluent quality requirements. Aeration alone accounts for 40 to 60 percent of a treatment plant’s total energy use. When 20 to 40 percent of influent flow is extraneous water — a common range for aging collection systems — the utility is spending a significant fraction of its energy budget to treat rainwater and groundwater to secondary treatment standards. For a southwestern Ontario municipality operating multiple wastewater treatment plants, the system-wide energy penalty from I&I could be $50,000 per year or$ 500,000 per year. Without measurement, the answer is a shrug.

The traditional approach to I&I investigation is well established: smoke testing, dye testing, CCTV inspection, and engineering judgment. These methods are valuable. They identify specific defects — cracked pipes, deteriorated joints, cross-connections — and they guide point repairs. But they are reactive, spatially limited, and expensive to deploy across an entire collection system. Smoke testing a single sub-catchment might cost $20,000 to$ 50,000 and take weeks. Scaling that across a multi-plant municipality means years of investigation before the full picture emerges.

What is missing is a system-wide, data-driven method to detect I&I severity at every plant, rank them by energy cost, and direct rehabilitation budgets where they will deliver the most return — all before a single smoke canister is deployed. The data for this analysis already exists. SCADA systems log influent flow at every plant. Environment Canada publishes hourly rainfall and temperature data for free. Monthly electricity bills quantify the energy cost. The analytical challenge is extracting the I&I signal from these routine data streams and translating it into actionable capital planning information.

This post presents four complementary detection methods — cross-correlation lag analysis, hydrograph decomposition, conductivity-based dilution tracing, and machine learning — demonstrated on synthetic data representing a multi-plant collection system with distinct I&I severity profiles. The methods are designed to work together: each provides an independent estimate of I&I severity, and agreement across methods builds confidence in the results. The final output is a ranked prioritization table that quantifies the energy penalty at each plant and evaluates the cost-benefit ratio of remediation scenarios.

2. Inflow vs. Infiltration — Two Failure Modes, Two Signatures

Inflow and infiltration enter sanitary sewers through fundamentally different mechanisms, and the distinction matters because each produces a different flow signature that requires a different detection approach.

Inflow is the rapid entry of surface water into the sanitary sewer system during or immediately after precipitation events. Sources include roof drain connections to sanitary laterals, sump pumps discharging to the sewer rather than to grade, deteriorated or missing manhole covers that allow surface water entry, and cross-connections between storm and sanitary systems. Inflow responds fast — minutes to hours after rainfall begins, the treatment plant sees a flow increase. The response is sharp, high-amplitude, and short-duration. When the rain stops, inflow subsides quickly.

Infiltration is the slow, sustained entry of groundwater into the sanitary sewer through structural defects in the pipe network. Sources include cracked pipes, deteriorated pipe joints, root intrusion pathways that have opened gaps in the pipe wall, and lateral connections where differential settlement has pulled joints apart. Infiltration responds slowly — hours to days after precipitation, as rainfall percolates through the soil and raises the local water table. The response is broad, lower-amplitude, and sustained. Infiltration can persist for days or weeks after a significant rain event, and in areas with chronically high water tables, it continues as a constant baseflow contribution even during dry weather.

Snowmelt is a third I&I driver — and in Ontario and other northern climates, it is often the worst. Spring freshet produces both rapid inflow from surface meltwater and sustained infiltration from saturated, thawing ground. The soil profile, frozen during winter, begins thawing from the surface down. Meltwater that cannot percolate through still-frozen deeper layers runs overland to manhole covers and surface connections (inflow), while meltwater that enters the thawed shallow soil zone finds every crack and joint in the pipe network (infiltration). The result is a prolonged I&I period — often weeks — that can exceed the peak response from any individual rain event.

A degree-day snowmelt model estimates the daily melt rate from temperature data:

$M = C_m \times \max(0, T - T_b)$

where $M$ is the daily snowmelt depth (mm/day), $C_m$ is the melt coefficient (typically 2 to 5 mm per degree-C per day, varying with land cover and solar exposure), and $T_b$ is the base temperature (0 degrees C). This is a simplified model — energy balance methods that incorporate solar radiation and wind speed are more accurate — but degree-day estimation has the practical advantage of requiring only temperature data, which Environment Canada provides hourly at no cost.

The combined framework is RDII — rainfall-derived inflow and infiltration — defined as the total flow at the treatment plant minus the dry weather flow baseline. RDII captures both the fast inflow response and the slow infiltration response in a single metric, and the physical response timescale is the diagnostic key: fast response indicates inflow dominance, slow response indicates infiltration dominance, and mixed response timescales indicate both mechanisms are active.

One important assumption underlies this entire methodology: the analysis assumes separated sanitary sewer systems where stormwater is collected in a separate storm drainage network. Combined sewer areas — where stormwater is intentionally collected in the same pipes as sanitary flow — require fundamentally different analysis because the “extraneous” flow is by design, not by defect. Combined sewer overflow (CSO) analysis is a distinct engineering discipline and is outside the scope of this framework.

Freeze-thaw cycles compound the problem over time. Seasonal ground movement from frost heave and thaw settlement progressively deteriorates pipe joints, particularly at service lateral connections. Each winter-spring cycle opens joints slightly wider, increasing the infiltration pathway for the following year. This cumulative degradation means I&I severity at a given location tends to increase over time in cold climates, making early detection and prioritization more valuable.

3. Method 1: Cross-Correlation Lag Analysis

The cross-correlation function (CCF) between rainfall intensity and sewer flow rate reveals the dominant time delay between precipitation at the surface and the flow response at the treatment plant. This lag time is the primary diagnostic for classifying I&I type — it directly measures how fast extraneous water moves from the ground surface into the sewer and through the collection system to the plant.

The CCF at lag $k$ is defined as:

$r_{xy}(k) = \frac{\sum_{t=1}^{N-k}(x_t - \bar{x})(y_{t+k} - \bar{y})}{\sqrt{\sum(x_t - \bar{x})^2 \cdot \sum(y_t - \bar{y})^2}} \tag{1}$

where $x_t$ is the rainfall (or snowmelt) intensity at time $t$ , $y_{t+k}$ is the sewer flow at time $t+k$ , and $k$ is the lag in time steps. The CCF value $r_{xy}(k)$ ranges from -1 to +1 and measures the linear correlation between the input signal and the output signal shifted by $k$ time steps. A peak in the CCF at lag $k^*$ indicates that the strongest linear relationship between rainfall and flow occurs with a delay of $k^*$ time steps.

Pre-whitening is essential. Both rainfall and sewer flow time series contain strong autocorrelation — flow at time $t$ is correlated with flow at time $t-1$ simply because pipe systems have hydraulic memory, and rainfall events have temporal structure. Computing the CCF directly on the raw time series produces inflated and misleading correlation values because the autocorrelation inflates the apparent cross-correlation. The standard remedy is pre-whitening: fit an ARIMA model to each time series independently, extract the residuals (the part of each series not explained by its own history), and compute the CCF on the residual series. The residuals are approximately white noise by construction, and the CCF computed on them reflects genuine cross-correlation between rainfall and flow rather than artifacts of internal autocorrelation.

The pre-whitening step also sharpens the CCF peak. Raw CCF plots often show broad, ambiguous peaks that make lag identification uncertain. Pre-whitened CCF plots produce narrower, more distinct peaks that permit more confident classification.

Interpreting the CCF peak is straightforward once the dominant lag is identified. A peak at lag 0 to 6 hours indicates inflow dominance — surface water entering the sewer rapidly through direct connections. A peak at lag 12 to 72 hours indicates infiltration dominance — groundwater entering slowly through pipe defects after the water table responds to precipitation. Multiple peaks at different lags indicate mixed I&I — both inflow and infiltration pathways are active. No significant peak at any lag suggests the plant has low I&I and is not rainfall-responsive.

The shape of the CCF peak matters as much as its location. A sharp, narrow peak indicates a well-defined response pathway. A broad, diffuse peak indicates multiple pathways with varying travel times — common in large service areas where different sub-catchments contribute I&I at different rates.

The CCF should be computed for both rainfall-flow and snowmelt-flow pairs. In many Ontario systems, the snowmelt-flow CCF reveals a different lag signature than the rainfall-flow CCF — spring melt often produces longer lags because the infiltration pathway through partially frozen ground differs from the pathway through unfrozen summer soil. Comparing rain-season and snowmelt-season CCF signatures provides additional diagnostic information about the physical I&I mechanisms and can reveal seasonal lag shifts attributable to freeze-thaw effects on pipe joints.

4. Method 2: Hydrograph Decomposition (DWF + RDII Separation)

Hydrograph decomposition separates total influent flow into its two components: the dry weather flow (DWF) baseline that represents legitimate sanitary flow, and the rainfall-derived inflow and infiltration (RDII) that represents extraneous water. This separation is the foundation for quantifying how much extraneous flow each plant receives and how much energy that flow consumes.

The definition is direct:

$Q_{\text{RDII}}(t) = Q_{\text{total}}(t) - Q_{\text{DWF}}(t) \tag{2}$

The DWF baseline must be established from periods when no precipitation influence is present. The standard practice is to identify dry weather periods — typically defined as 72 or more consecutive hours with no rainfall and no active snowmelt — and compute the median diurnal flow pattern from these periods. The median is preferred over the mean because it is robust to outliers from industrial discharges, sensor glitches, or residual wet-weather tails. The DWF baseline should be computed separately for weekdays and weekends (commercial and institutional flows reduce on weekends) and should incorporate seasonal variation (groundwater-driven baseflow infiltration is typically higher in spring and lower in late summer).

Once the DWF baseline is established, RDII at any time step is the positive difference between observed flow and DWF. Negative differences (observed flow below DWF) are set to zero — they represent normal DWF variability, not negative I&I.

The RDII response to a precipitation event can be modeled as a convolution of the rainfall input with a unit hydrograph — a transfer function that describes how the collection system transforms a unit pulse of rainfall into a flow response at the plant:

$Q_{\text{RDII}}(t) = \sum_{\tau=0}^{T} P(\tau) \cdot h(t - \tau) \tag{3}$

where $P(\tau)$ is the rainfall intensity at time $\tau$ and $h(t - \tau)$ is the unit hydrograph value at lag $t - \tau$ . The convolution captures the distributed, time-delayed nature of the I&I response: each increment of rainfall produces a flow response that is spread over time according to the shape of the unit hydrograph, and the total RDII at any time is the superposition of all these individual responses.

The EPA SSOAP (Sanitary Sewer Overflow Analysis and Planning) toolbox formalizes this with the RTK unit hydrograph model, which uses three triangular unit hydrograph components to represent three distinct I&I pathways:

Fast (R1, T1, K1): Represents inflow — rapid surface water entry. Short time to peak (T1, typically 1 to 4 hours), sharp recession (K1), and a volume fraction R1 that captures the proportion of total RDII arriving through fast pathways.
Medium (R2, T2, K2): Represents fast infiltration — water entering through near-surface defects. Intermediate time to peak (T2, typically 4 to 12 hours) and a broader recession.
Slow (R3, T3, K3): Represents slow infiltration — groundwater-driven entry through deeper pipe defects. Long time to peak (T3, typically 12 to 72 hours), very broad recession, and often the largest volume contribution in systems with significant groundwater infiltration.

The R values (dimensionless fractions) sum to the total runoff coefficient — the fraction of rainfall over the service area that enters the sewer as RDII. Typical total R values for separated sanitary sewers range from 0.01 to 0.10, meaning 1 to 10 percent of rainfall volume enters the sewer. Values exceeding 0.10 indicate severe I&I; values below 0.01 suggest a tight system.

An alternative decomposition approach uses STL (Seasonal and Trend decomposition using Loess) to separate the flow time series into three additive components: trend (long-term changes in baseflow infiltration), seasonal (the repeating 24-hour diurnal pattern of sanitary flow), and remainder (wet-weather response plus anomalies). The STL remainder component during wet weather periods approximates the RDII signal. STL has the advantage of being a purely statistical method that requires no hydrological assumptions, but it cannot distinguish between inflow and infiltration components the way the RTK model can.

5. Method 3: Dilution Tracers (Conductivity and Temperature)

Conductivity provides a direct, physics-based measurement of dilution that is independent of the flow-based methods described in Sections 3 and 4. Sanitary wastewater carries dissolved salts, detergents, and organic compounds that produce a characteristic electrical conductivity — typically 1,000 to 2,000 microsiemens per centimetre (uS/cm). Stormwater and shallow groundwater have much lower conductivity — typically 100 to 500 uS/cm, depending on local soil chemistry. When extraneous water enters the sanitary sewer, it dilutes the wastewater, and the influent conductivity at the treatment plant drops in proportion to the volume fraction of extraneous flow.

The dilution fraction is calculated from a simple mixing model:

$f_{\text{extraneous}} = \frac{\sigma_{\text{DWF}} - \sigma_{\text{observed}}}{\sigma_{\text{DWF}} - \sigma_{\text{extraneous}}} \tag{4}$

where $\sigma_{\text{DWF}}$ is the baseline dry weather conductivity (established from the same dry weather periods used for DWF flow estimation), $\sigma_{\text{observed}}$ is the measured conductivity during wet weather, and $\sigma_{\text{extraneous}}$ is the conductivity of the extraneous water source. The result $f_{\text{extraneous}}$ is the fraction of total influent flow that is extraneous — a value that can be compared directly against the RDII fraction computed from flow decomposition.

The conductivity of the extraneous water source requires an assumption or measurement. Stormwater conductivity varies with antecedent dry period (longer dry periods produce higher first-flush conductivity from accumulated surface deposits), season (road salt increases winter stormwater conductivity significantly in Ontario), and location. Groundwater conductivity depends on local geology and aquifer chemistry. A representative value of 200 to 400 uS/cm is commonly used for stormwater in the absence of direct measurement, but winter values in road-salted areas can exceed 1,000 uS/cm, which reduces the conductivity contrast and makes dilution detection less sensitive. Road salt contamination of infiltrating groundwater can similarly confound the method during late winter and early spring.

Temperature serves as a complementary tracer. Sanitary wastewater arriving at the plant typically ranges from 15 to 25 degrees C depending on season — warmer than ambient in winter, often similar to ambient in summer. Stormwater and shallow groundwater are closer to ambient ground or air temperature. During a summer rain event, cold stormwater entering the sewer produces a detectable temperature drop at the plant. During winter, snowmelt entering the sewer is near 0 degrees C and produces a sharp temperature drop against the warmer wastewater background.

The advantage of conductivity and temperature as tracers is their independence from the flow measurement. Flow-based methods (CCF, hydrograph decomposition) require accurate flow metering and can be confounded by sensor drift, fouling, or calibration errors. Conductivity and temperature provide a separate line of evidence that either confirms or contradicts the flow-based I&I estimate. When the two approaches agree, confidence in the result increases. When they disagree, the discrepancy points to a data quality issue that warrants investigation.

Continuous conductivity sensors suitable for sewer environments cost $200 to$ 500 per installation and can be integrated into existing SCADA systems. Not all plants have conductivity instrumentation — this is a “nice-to-have” method rather than a required one. But where conductivity data is available, it provides a valuable cross-check that strengthens the overall assessment.

6. Method 4: Machine Learning for RDII Prediction

The three methods described above — CCF lag analysis, hydrograph decomposition, and dilution tracing — are interpretable, physics-informed approaches. Machine learning adds a fourth layer: the ability to learn complex, nonlinear relationships between weather inputs and RDII response that the analytical methods may not fully capture. The ML approach does not replace the physics-based methods — it complements them by providing an independent prediction that can be compared for consistency.

The RDII prediction model takes the general form:

$\hat{Q}_{\text{RDII}}(t) = f\left(P_{t-1}, \ldots, P_{t-n},\; M_{t},\; Q_{t-1},\; \sigma_t,\; T_t,\; \text{GWL}_t\right) \tag{5}$

where $P_{t-1}, \ldots, P_{t-n}$ are lagged rainfall values (capturing antecedent precipitation), $M_t$ is the degree-day snowmelt estimate, $Q_{t-1}$ is the previous flow (capturing system memory), $\sigma_t$ is conductivity if available, $T_t$ is temperature, and $\text{GWL}_t$ is groundwater level if available. Not all inputs are required — the model adapts to available data. At minimum, lagged rainfall, temperature (for snowmelt estimation), and previous flow provide a functional feature set.

Three ML architectures are well-suited to this problem, each with distinct strengths.

LSTM (Long Short-Term Memory) networks excel at capturing temporal dependencies in sequential data. The rainfall-to-RDII relationship is inherently sequential — today’s infiltration depends on yesterday’s rainfall, last week’s rainfall, and the cumulative antecedent moisture condition. LSTM networks learn these multi-scale temporal dependencies from data without requiring explicit specification of lag structures. A 2024 IWA study demonstrated that LSTM networks outperform traditional RTK unit hydrograph models for RDII prediction, particularly during complex, multi-event sequences where antecedent moisture effects are significant.

XGBoost (gradient-boosted decision trees) provides interpretable feature importance that directly answers a practical question: what drives I&I at this plant? SHAP (SHapley Additive exPlanations) values from an XGBoost model reveal whether a plant’s RDII is primarily driven by recent rainfall, antecedent moisture, snowmelt, or groundwater level. This feature importance ranking has direct operational value — a plant where RDII is dominated by short-lag rainfall features is inflow-dominated, while a plant where antecedent precipitation index and groundwater features dominate is infiltration-driven. XGBoost also handles missing features gracefully, which matters when conductivity or groundwater data is not available for all plants.

Autoencoder-based anomaly detection addresses a different question: which time periods are I&I-impacted? An autoencoder trained exclusively on dry weather flow data learns the normal diurnal and seasonal patterns. When applied to the full time series, periods where the reconstruction error exceeds a threshold are flagged as anomalous — and in a sewer context, flow anomalies during or after precipitation are predominantly I&I. This classification approach does not predict RDII magnitude, but it provides a binary detection that identifies exactly when I&I events begin and end, which is useful for event-based RDII volume accounting.

Model validation follows standard time-series practice: a chronological train-test split (never random — temporal leakage invalidates the results). Two years of data for training, one year for testing, with the test year containing both rain-season and snowmelt-season events. Performance metrics include R-squared, RMSE, and for the anomaly detection approach, precision and recall for I&I period classification. Prediction-versus-actual scatter plots and residual analysis by season provide visual confirmation that the model generalizes across weather conditions.

The ML models do not need to outperform the analytical methods to be useful. Their primary value is as an independent check — if the XGBoost model, the CCF analysis, and the hydrograph decomposition all agree that Plant C has the worst I&I, the confidence in that conclusion is much higher than if a single method produced it.

7. Data Requirements — Keeping It Achievable

A common failure mode in infrastructure analytics is specifying data requirements that are theoretically ideal but practically impossible. Detailed pipe condition data, comprehensive GIS inventories, and continuous groundwater monitoring are valuable — but most utilities do not have them, and requiring them as prerequisites blocks any analysis from starting. This methodology is deliberately designed around data that most municipal wastewater utilities already collect or can obtain at no cost.

Tier 1 — Required:

SCADA influent flow. Every wastewater treatment plant logs influent flow. Most SCADA systems record at 5-minute to 15-minute intervals. Fifteen-minute data is adequate for infiltration detection (slow response). Five-minute data is preferred for inflow detection (fast response), but 15-minute data will still capture the dominant inflow signal — the CCF peak will be broader but still identifiable. Two to three years of continuous data provides sufficient wet weather events and snowmelt seasons for robust analysis.
Environment Canada rainfall and temperature. Hourly precipitation and temperature data from Environment Canada weather stations is freely available at climate.weather.gc.ca. This is the same data used for regulatory reporting at many utilities. Temperature data is essential — not just for snowmelt estimation via the degree-day model, but for identifying freeze-thaw periods that affect CCF lag interpretation. Select the weather station closest to the service area. For service areas larger than 50 square kilometres, rainfall spatial variability may warrant using multiple stations, but a single representative station is adequate for plant-level screening.
Monthly energy consumption. Plant-level electricity bills or SCADA power metering data, at monthly resolution. This translates RDII volumes into energy costs and enables the cost-benefit analysis in Section 8. Monthly resolution is sufficient because the energy penalty calculation uses annual RDII volume, not hourly power demand.

Tier 2 — Optional (enhances analysis):

Influent conductivity. Enables Method 3 (dilution tracing) as an independent cross-check. Available at plants with conductivity instrumentation on the influent channel.
Influent temperature. Complementary tracer for dilution detection, particularly useful during snowmelt when the temperature contrast between meltwater and sanitary wastewater is large.
Intermediate pump station flow data. Where SCADA data exists for pump stations within the collection system, the analysis can extend beyond plant-level ranking to sub-catchment spatial resolution. This identifies which zones within a plant’s service area contribute disproportionate I&I — valuable for targeting rehabilitation investments to the highest-impact locations.

Explicitly NOT required:

Pipe age, material, or condition data
GIS mapping of the collection system
Manhole survey records
Asset management database
CCTV inspection reports
Groundwater monitoring wells

These data sources are valuable for remediation planning — deciding what to fix once high-I&I areas are identified. But they are not needed for the detection and ranking phase. The methodology ranks plants by observed I&I behavior from SCADA data alone, using the flow response to weather events as the diagnostic signal. This keeps the barrier to entry low: any utility with SCADA and an internet connection can begin the analysis.

The separated sanitary sewer assumption applies to all data interpretation. Plants receiving flow from combined sewer areas require different treatment — the “extraneous” flow is by design, not by defect, and the CCF and RDII methods will produce misleading results if applied without adjustment.

8. Quantifying the Energy Penalty

The detection methods in Sections 3 through 6 identify and quantify I&I. This section translates those volumes into energy costs — the metric that converts a collection system problem into a capital planning argument.

Wastewater treatment energy intensity varies with facility size, process type, and effluent quality requirements. National averages for secondary treatment fall between 1,000 and 3,000 kWh per million gallons (MG), with larger plants at the lower end of the range (economies of scale in aeration and pumping) and smaller plants at the upper end. For a given plant, the energy intensity can be calculated directly from SCADA flow totals and utility bills.

The energy penalty from I&I concentrates in two process areas. Pumping energy scales approximately linearly with flow volume — every additional gallon of I&I must be lifted through the collection system and into the plant. Aeration energy scales with both flow and load — but since I&I dilutes the influent without adding proportional organic load, aeration systems that control to a dissolved oxygen setpoint will consume similar blower energy at higher flow because the same oxygen transfer must occur in a larger volume of water passing through the aeration basin. The marginal energy cost of I&I flow is therefore somewhat lower than the average treatment energy intensity, but for screening purposes, using the plant’s average energy intensity provides a reasonable and conservative estimate.

The annual energy penalty is:

$\boxed{E_{\text{penalty}} = Q_{\text{I\&I}} \times \text{EI}_{\text{treatment}} \times 365} \tag{6}$

where $Q_{\text{I\&I}}$ is the average daily I&I volume (MG/day), $\text{EI}_{\text{treatment}}$ is the plant’s energy intensity (kWh/MG), and the factor of 365 converts to annual energy. The dollar cost follows by multiplying by the blended electricity rate.

Consider a plant treating an average of 5 MGD with an I&I fraction of 30 percent (1.5 MGD of extraneous flow) and an energy intensity of 1,800 kWh/MG. The annual energy penalty is 1.5 multiplied by 1,800 multiplied by 365, which equals approximately 985,500 kWh per year. At $0.12 per kWh, that is$ 118,000 per year — spent treating water that should never have entered the sewer.

The cost-benefit ratio for I&I rehabilitation compares the present value of avoided energy costs against the capital cost of pipe repairs:

$\boxed{\text{BCR} = \frac{\Delta E_{\text{annual}} \times C_{\text{energy}} \times L_{\text{asset}}}{ C_{\text{rehabilitation}}}} \tag{7}$

where $\Delta E_{\text{annual}}$ is the annual energy savings from I&I reduction (kWh/year), $C_{\text{energy}}$ is the electricity cost (dollars per kWh), $L_{\text{asset}}$ is the expected useful life of the rehabilitation (years — typically 50 years for CIPP lining, 30 years for manhole rehabilitation), and $C_{\text{rehabilitation}}$ is the capital cost of the intervention. A BCR greater than 1.0 indicates that the energy savings alone justify the rehabilitation cost over the asset life, before considering other benefits such as reduced SSO risk, deferred capacity expansion, and reduced chemical costs.

The BCR calculation intentionally uses only energy savings as the numerator benefit — it does not include avoided capacity expansion, reduced chemical costs, or SSO risk reduction. This is conservative by design: if the BCR exceeds 1.0 on energy alone, the rehabilitation is justified even if the other benefits are difficult to quantify. Where BCR falls between 0.5 and 1.0, the full suite of non-energy benefits may still justify the investment, but the case requires a broader analysis.

9. Verifying the Analysis — How to Know Your Results Are Right

Running the numbers is not enough. Every detection method produces output, but output without verification is not actionable — it is a liability. This section presents concrete quality checks at each stage of the methodology. An analyst who cannot pass these checks should not present results to decision-makers.

DWF baseline sanity checks. The dry weather flow diurnal pattern should show a recognizable shape: a morning peak (06:00 to 09:00) from residential wake-up activity, a secondary peak in the evening (17:00 to 21:00), and an overnight minimum (01:00 to 05:00) at roughly 30 to 50 percent of the daytime peak. If the diurnal pattern is flat, the flow meter may be damped or averaging excessively. If there is no overnight minimum, baseflow infiltration is likely already elevated. Compare the per-capita DWF against the Ontario MECP design guideline of approximately 250 litres per capita per day for sanitary flow. If the DWF per-capita rate significantly exceeds this value, chronic groundwater infiltration is contributing to the baseline even during dry weather. Cross-plant comparison within the municipality provides additional context — if one plant’s per-capita DWF is double the others, that plant has a baseline infiltration problem.

RDII mass balance. The total RDII volume from each precipitation event should be physically plausible relative to the rainfall volume falling on the service area. The runoff coefficient — RDII volume divided by rainfall volume over the service area — should fall between 0.01 and 0.10 for separated sanitary sewers. A coefficient of 0.02 means 2 percent of rainfall enters the sewer; a coefficient of 0.08 means 8 percent. If the computed runoff coefficient exceeds 0.10, the RDII estimate is likely inflated by DWF baseline error, flow meter inaccuracy, or incorrect service area delineation. If the RDII volume exceeds the total rainfall volume — a coefficient greater than 1.0 — something is fundamentally wrong with the data or the calculation.

CCF lag plausibility. Inflow lags should fall between 0 and 6 hours. Infiltration lags should fall between 12 and 72 hours. Lags outside these ranges warrant investigation — a computed lag of 200 hours, for example, likely reflects autocorrelation artifacts from inadequate pre-whitening rather than a genuine 8-day travel time. If no significant CCF peak exists at any lag, the plant may have genuinely low I&I (which is a valid finding), or the rainfall station may be too distant to represent precipitation over the service area.

Conductivity cross-check. If conductivity data is available, the flow-based RDII fraction and the conductivity-based dilution fraction should be broadly consistent. If flow increases by 40 percent during wet weather but conductivity does not drop, the flow increase may not be extraneous water — it could be a flow meter calibration shift triggered by high water level in the influent channel. Conversely, if conductivity drops sharply but flow does not increase proportionally, the conductivity sensor may be responding to a local mixing effect rather than a plant-wide dilution. Persistent disagreement between flow and conductivity signals is a data quality flag, not a methodological failure.

ML model validation. The train-test split must be chronological — never random. Random splitting allows the model to learn future weather patterns during training, producing artificially high performance that will not generalize. The test set should include at least one full rain season and one full snowmelt season. Scatter plots of predicted versus actual RDII should cluster along the 1:1 line without systematic bias. Residual plots should show no seasonal pattern — if residuals are consistently positive during spring snowmelt, the model is underestimating melt-driven I&I and the snowmelt features need refinement. Report R-squared, RMSE, and for anomaly detection, precision and recall.

Multi-method consistency. The four methods provide independent I&I severity estimates. If the CCF analysis ranks Plant A as severe, hydrograph decomposition ranks Plant A as severe, and XGBoost feature importance shows Plant A is highly rainfall-responsive — the confidence in that assessment is high. If two methods agree and one disagrees, investigate the outlier method’s assumptions and data inputs before dismissing it.

Snowmelt cross-check. The degree-day snowmelt estimate should correlate with observed spring flow increases. Compare melt-driven RDII timing against temperature records. If flow spikes occur during periods when the temperature is consistently below freezing and the snowmelt model predicts zero melt, the flow increase is not melt-driven — it may be a freeze-thaw effect on pipe joints or a groundwater response to upstream melt. Adjust the melt coefficient $C_m$ if the timing and magnitude of predicted melt do not align with observed flow increases.

Seasonal repeatability. A single storm analysis is unreliable. The I&I signature should repeat consistently across multiple wet weather events within a season and across multiple seasons. Compare rain-season lag signatures against snowmelt-season lag signatures — if the dominant lag shifts by 12 hours or more between seasons, freeze-thaw effects on pipe joints or seasonal groundwater table variation may be altering the I&I pathway. Consistent signatures across two or more years of data provide the strongest basis for plant ranking.

10. Mitigation — Operational and Capital Measures

Detection and quantification are not endpoints — they are inputs to remediation planning. This section organizes mitigation measures by type, from investigation to capital rehabilitation to green infrastructure, with cost and effectiveness context.

Operational investigation confirms and localizes I&I before capital is committed. Smoke testing introduces non-toxic smoke into sewer segments to identify surface connections — smoke emerging from roof drains, yard drains, or foundation vents confirms inflow pathways. Dye testing traces specific connections by introducing colored dye into suspected sources and monitoring for its appearance in the sewer. CCTV inspection using NASSCO PACP (Pipeline Assessment Certification Program) protocol provides internal pipe condition assessment — cracks, joint separations, root intrusion, and lateral connection defects are graded by severity. Temporary flow monitoring with portable area-velocity meters in manholes provides sub-catchment flow data that localizes I&I to specific pipe segments. These investigation methods are targeted and sequential: the data-driven plant ranking from Sections 3 through 6 directs investigation to the highest-priority service areas, and the investigation methods identify the specific defects within those areas.

Capital rehabilitation addresses the physical defects. CIPP (cured-in-place pipe) lining is the most common trenchless rehabilitation method for mainline sewers. A resin-impregnated liner is inserted into the existing pipe and cured in place, sealing cracks, joint separations, and infiltration pathways without excavation. Published I&I reduction rates from CIPP range from 10 to 30 percent — the wide range reflects variation in pre-rehabilitation defect severity and groundwater conditions. Pipe bursting replaces the existing pipe with a new one by pulling an expander head through the old pipe while simultaneously pulling in a new HDPE pipe. Manhole rehabilitation — typically epoxy or polyurethane coating of the interior surfaces — seals frame-chimney joints and wall infiltration at a lower cost per unit than mainline rehabilitation.

Private lateral programs address the single largest source of I&I in many systems. WERF (Water Environment Research Foundation) studies consistently show that private service laterals — the pipes connecting individual buildings to the sewer main — account for 50 percent or more of total I&I in separated sanitary systems. These laterals are typically on private property and outside the utility’s direct maintenance responsibility, creating a challenging governance and funding problem. The Town of Fort Erie, Ontario, implemented a successful extraneous flow reduction program offering incentives up to $1,500 per property for lateral replacement or foundation drain disconnection. Programs like this require municipal council approval and public outreach but address the largest single I&I source that capital rehabilitation of mains cannot reach.

Green infrastructure reduces I&I indirectly by reducing the volume of stormwater reaching the ground surface near sewer infrastructure. Downspout disconnection programs redirect roof drainage from foundation drains (which may be connected to the sanitary sewer) to grade. Rain gardens and bioretention cells infiltrate stormwater on-site, reducing the soil saturation that drives groundwater infiltration. Permeable pavement reduces surface runoff that enters manholes and surface connections. These measures reduce I&I and provide stormwater management co-benefits, but their I&I reduction is difficult to quantify precisely and depends heavily on proximity to sewer defects.

Flow equalization does not reduce I&I but manages its consequences. Equalization basins or in-system storage attenuate peak wet-weather flows, preventing hydraulic overload at the treatment plant and reducing the risk of sanitary sewer overflows (SSOs). Equalization is a capacity management tool, not an I&I reduction tool — the extraneous water still enters the system and still consumes treatment energy, but the peak is spread over a longer period.

Ontario regulatory context frames the cost-effectiveness threshold. The MECP (Ministry of the Environment, Conservation and Parks) Design Guidelines for Sewage Works (2008) establish an infiltration threshold of 0.14 L/(mm of pipe diameter per day) per metre of pipe length. Below this threshold, rehabilitation is generally not considered cost-effective. Environmental Compliance Approvals (ECAs) under the Ontario Water Resources Act govern treatment plant capacity, and I&I-driven capacity exceedances can trigger regulatory action. The F-5-1 guideline provides the compliance framework for determining acceptable effluent quality during wet weather events.

11. Prioritization Framework — Ranking Plants for Remediation

Detection and quantification produce a dataset. Prioritization translates that dataset into a capital planning recommendation. For a municipality operating multiple treatment plants with limited rehabilitation budgets, the question is not whether to address I&I but where to address it first for maximum return.

The prioritization framework builds a composite I&I severity score for each plant from the SCADA-derived metrics produced by the detection methods. No asset condition data is required — the ranking is based entirely on observed I&I behavior and its energy cost.

Five metrics contribute to the composite score, each capturing a different dimension of I&I severity:

RDII volume fraction (weight: 0.30): The fraction of annual influent flow attributable to I&I, computed from hydrograph decomposition. A higher fraction means more extraneous flow relative to legitimate sanitary flow.
Energy cost per capita (weight: 0.25): The annual energy penalty (Equation 6) divided by the served population. This normalizes for plant size and identifies plants where I&I imposes the highest per-person energy burden.
CCF peak magnitude (weight: 0.20): The maximum pre-whitened CCF value between rainfall (or snowmelt) and flow. A higher peak indicates a stronger rainfall-to-flow coupling and more rainfall-responsive I&I.
Snowmelt ratio (weight: 0.10): The ratio of snowmelt-season RDII volume to rain-season RDII volume. Plants where snowmelt-driven I&I dominates may require different remediation strategies (e.g., foundation drain disconnection) than plants where rain-driven inflow dominates.
BCR (weight: 0.15): The cost-benefit ratio from Equation 7 for a representative rehabilitation scenario (e.g., CIPP lining of the worst-condition mainline segments). Higher BCR indicates higher remediation ROI.

The weights reflect a balance between I&I severity (what is happening), energy impact (what it costs), and economic return (what rehabilitation would save). The weights can be adjusted to reflect local priorities — a municipality under regulatory pressure for SSO reduction might increase the weight on RDII volume fraction, while a municipality focused on energy cost reduction might increase the weight on energy cost per capita.

Plants are ranked by composite score in descending order. The top-ranked plants represent the highest-priority candidates for detailed investigation (smoke testing, CCTV, flow monitoring) and rehabilitation design. The ranked list provides a defensible basis for capital improvement plan submissions and multi-year rehabilitation program design.

Where intermediate pump station flow data is available, the analysis extends below the plant level to sub-catchment spatial resolution. Each pump station’s service area can be analyzed independently using the same CCF and RDII methods applied at the plant level. This identifies which zones within a plant’s service area contribute disproportionate I&I — enabling targeted rehabilitation in the highest-impact sub-catchments rather than system-wide programs that spend capital in areas with low I&I. Sub-catchment prioritization multiplies the value of the plant-level ranking by concentrating investigation and rehabilitation dollars within the highest-priority service areas at the highest-priority plants.

The prioritization framework produces an actionable output for a municipality with constrained capital budgets: a ranked list of plants (and, where data permits, sub-catchments) ordered by composite I&I severity and remediation ROI, built entirely from data already being collected, requiring no asset condition surveys or pipe inspections to generate.

12. Limitations and Practical Considerations

Every analytical framework rests on assumptions, and this one is no exception. Understanding where those assumptions hold and where they introduce uncertainty is essential for interpreting results correctly and communicating findings honestly to decision-makers.

Sensor accuracy in sewer environments is a persistent challenge. Influent flow meters in wastewater plants operate in a harsh environment — submerged in gritty, corrosive wastewater with floating debris. Ultrasonic and electromagnetic flow meters require regular calibration and can drift between maintenance intervals. A 5 percent flow measurement error translates directly to a 5 percent error in RDII estimation. Redundant flow measurement (primary element plus check meter) improves confidence but is not available at every plant. The verification checks in Section 9 — particularly the per-capita DWF comparison and cross-plant consistency — help identify plants where flow measurement quality is suspect.

Rainfall spatial variability is the most significant environmental uncertainty. A single rain gauge or weather station represents precipitation at one point. Convective summer storms can produce intense rainfall over one sub-catchment while an adjacent sub-catchment 5 kilometres away receives nothing. If the weather station is in the dry zone and the treatment plant sees a flow spike, the CCF analysis will fail to detect the correlation. For plant-level screening, a single representative weather station is usually adequate because convective events average out over a multi-year analysis period. For sub-catchment analysis using pump station data, rainfall spatial resolution becomes more critical and may require multiple gauges or radar-based precipitation estimates.

Antecedent moisture conditions affect infiltration response. The same rainfall event produces more I&I when the soil is already saturated (spring, after prolonged wet weather) than when the soil is dry (late summer, after a drought). The ML models in Section 6 capture this through the antecedent precipitation index feature, but the analytical methods (CCF, hydrograph decomposition) treat each event independently. Seasonal stratification — computing separate CCF and RDII statistics for spring, summer, and fall — partially addresses this limitation.

Seasonal groundwater effects create a time-varying infiltration baseline. In many Ontario systems, the water table rises in spring from snowmelt recharge and falls through summer from evapotranspiration. This seasonal variation produces a slow, broad infiltration signal that may be classified as DWF variation rather than I&I. The DWF baseline estimation should account for seasonal variation by computing separate baselines for each season or by using a trend-aware decomposition method like STL.

Snowmelt estimation uncertainty is inherent in the degree-day method. The melt coefficient $C_m$ varies with land cover (forested vs. urban), solar exposure (south-facing slopes vs. north-facing), and wind speed. The simple degree-day model does not capture radiation-driven melt on sunny days below 0 degrees C or rain-on-snow events that produce rapid, large-volume melt. Energy balance snowmelt models are more accurate but require radiation and wind data that are not always available from nearby weather stations. For the screening purpose of this methodology, the degree-day approach provides adequate accuracy — the goal is to identify which plants respond most strongly to snowmelt, not to predict exact melt volumes.

The separated sanitary sewer assumption must be verified, not assumed. Many Ontario municipalities have areas where separated systems were constructed adjacent to or downstream of older combined sewer areas, or where partial separation was completed decades ago with undocumented cross-connections remaining. Applying the I&I detection methodology to a system receiving intentional stormwater flow will produce inflated I&I estimates and misleading plant rankings. Confirm the sewer system type for each plant’s service area before interpreting results.

Model drift affects the ML models over time. Collection system deterioration, land use changes, population growth, and climate shifts alter the rainfall-to-flow relationship. Models trained on historical data will degrade in predictive performance as conditions change. Annual retraining on the most recent two to three years of data maintains model currency. The analytical methods (CCF, hydrograph decomposition) are less susceptible to drift because they are recomputed from scratch on each analysis cycle rather than depending on previously learned parameters.

Freeze-thaw effects on pipe joints produce a seasonal shift in I&I characteristics that complicates year-round analysis. Ground movement from frost heave and thaw settlement opens and closes pipe joints cyclically, potentially altering the dominant lag signature between winter and spring. Comparing CCF results from the frost season against the non-frost season can reveal these shifts — a plant whose dominant lag shortens from 48 hours in winter to 12 hours in spring may be experiencing freeze-thaw joint opening that converts slow infiltration pathways into faster ones during the thaw period.

Regulatory framing is important for Ontario context. The MECP infiltration threshold of 0.14 L/(mm of pipe diameter per day) per metre of pipe length provides a benchmark, but it applies to pipe-level assessment rather than plant-level screening. The plant-level methodology presented here identifies where to investigate; pipe-level assessment using the MECP threshold determines which specific segments warrant rehabilitation. The two scales of analysis are complementary, not competing.