AI for Water Utility Pump Scheduling: Start with Service Constraints

Make the pumping plan answer to the service requirement

A pump schedule can look cheaper on a spreadsheet while leaving a water utility with insufficient reserve, an unsuitable pressure profile or poorly turned-over storage. A useful AI project therefore starts with the service conditions the utility must maintain.

The opportunity is specific: forecast demand, compare feasible pumping options and give operators a better-supported plan. Begin with one drinking-water pumping station, pressure zone and connected storage. A whole-network rollout can wait until the team understands the data and operational consequences.

EPA's energy-efficiency guidance recommends assessing energy use, establishing a baseline and understanding electricity costs. That is a sound starting point for an AI pilot, too. General energy-efficiency opportunities should not be presented as savings achieved by AI. [1]

Separate forecasting, optimization and hydraulic checks

These are different jobs and should remain visible in the workflow.

Forecasting estimates demand over the scheduling period. Machine learning may help identify recurring patterns and changes, but forecasts should include uncertainty and be compared with a simple baseline, such as comparable recent days.

Optimization proposes pump combinations and operating periods within the approved constraints. The objective might be energy consumption, electricity cost or a defined balance of cost and equipment duty. Name the objective so that everyone knows what the schedule is trying to improve.

Hydraulic and water-quality analysis tests the candidate against the network model and operating requirements. EPA's EPANET supports extended-period simulation of pressurized networks, including pressure, tank levels, pumping energy, chemical concentration and water age. It is engineering simulation software; using it does not by itself validate an AI recommendation. [2]

A language model can explain the resulting decision packet or retrieve an approved procedure. It should not invent pump curves, hydraulic results or operating limits.

Build a baseline that explains the bill

Monthly electricity bills can identify a cost problem. A schedule-level comparison needs time-aligned operational evidence.

For the selected boundary, assemble interval power, flow, pump status and speed, pressure, storage levels, equipment availability and tariff periods. Include the water-quality information the utility uses to assess the proposed change. Record clock alignment, units, sensor quality and missing values.

Ask a practical question before training a model: can the team reconstruct why yesterday's pumps ran when they did?

A maintenance restriction, a temporary valve configuration or an operator response may explain a pattern that an algorithm would otherwise label inefficient. Keep those events in the baseline. Where the reason is unknown, say so rather than assigning a convenient explanation.

Use the current schedule, a straightforward rule-based alternative and the proposed AI-assisted schedule as comparators. If the simpler method delivers the same operational benefit, the additional model may not be justified.

Put the acceptance envelope in writing

Operations and engineering should specify the conditions a candidate schedule must satisfy. Depending on the system, these may include pressure requirements, usable storage, emergency reserve, permitted pump operating ranges, start-stop limits, standby availability and water-quality objectives.

Pressure and storage turnover matter alongside energy. EPA's distribution-system resources discuss pressure management and explain how increased water age can contribute to declining disinfectant residual and other water-quality concerns. A simulated water-age value is a useful diagnostic, not a direct measurement of drinking-water safety. [3]

Treat these as explicit acceptance checks. The optimizer should not be able to exchange a required service condition for a lower predicted electricity bill.

Also define when the assistant must withhold a recommendation. Examples include stale tank-level data, an unavailable pump with uncertain status, an unmodelled network configuration or demand outside the evaluated range. Existing control logic, alarms, interlocks and emergency procedures remain in place.

Give the operator a small, inspectable decision packet

The operator should be able to see what changes and why without opening a large model report. Each recommendation should show:

• Proposed pump operating periods and the current-plan comparison
• Predicted electricity use and predicted cost as separate figures
• Starting and ending storage assumptions
• Demand range, input timestamps and relevant equipment restrictions
• The constraints closest to their limits
• Hydraulic-check results and unresolved exceptions
• Expiry time and the approved fallback

Keep rejected candidates available for engineering review. Their failure reasons may reveal that there is little flexibility in the current system, or that an equipment or data issue should be addressed before optimization.

For an initial deployment, the decision packet is advisory. Record the operator's acceptance, change or rejection and the reason. That feedback is part of the evaluation, not an inconvenience to remove.

Test the awkward days before changing schedules

Replay conditions that challenge a demand forecast and the network model: unusually high demand, prolonged low demand, one pump unavailable, a telemetry gap and a change in topology. Check both feasibility and the workflow's ability to stop when it lacks evidence.

Then run live in read-only mode. Compare proposed schedules with actual operating decisions while the existing process stays authoritative. Investigate disagreement with operators before assuming that a model has found an improvement.

Any controlled trial should use the utility's authorization and change process. Predefine monitoring, abort criteria and restoration of the prior operating plan. A pilot must remain easy to stop when field conditions depart from the assumptions.

Verify savings without borrowing from tomorrow's storage

Electricity cost and electricity consumption can move differently. Shifting pumping into a cheaper tariff period may reduce cost without reducing kilowatt-hours. Report both.

Likewise, a trial that ends with less water in storage can appear more efficient because it leaves replenishment for a later period. Compare starting and ending inventory, pumped volume, operating head, equipment configuration and service conditions.

Use energy per unit volume carefully. It is informative within a comparable boundary, but it can conceal a different lift, pressure duty or storage state. Document the baseline method and any adjustments rather than presenting one percentage as the entire result.

The business scorecard should include operator workload, service exceptions, rejected recommendations and fallback use alongside energy and cost. A schedule that requires constant intervention may be expensive even when its simulated bill is lower.

Adapt the method carefully for wastewater

A drinking-water storage strategy cannot simply be transferred to a sewage lift station. Wet-weather inflow, overflow margin, detention, solids handling and downstream treatment capacity create a different operating problem.

EPA's lift-station guidance describes the balance between wet-well storage, pump cycling and excessive detention. Those engineering considerations belong in any wastewater scheduling study; current site requirements and competent design review determine the actual limits. [4]

Start with one operating boundary, a credible baseline and an operator-reviewed recommendation. Expand only when measured results demonstrate useful flexibility under real service conditions.

See NeoBram's water and wastewater AI applications and industrial AI shadow-mode guide for the broader deployment context.