Choose the forecast horizon
A forecast for tomorrow’s pumping schedule answers a different question from a forecast for a growth corridor in 2046. State the spatial scale, time step, horizon and decision before selecting a method. Keep annual volume, average day, peak day and peak hour as separate quantities.
Build a transparent baseline
Reconcile production, transfers, storage changes and billed consumption over a common period. Separate residential and non-residential demand, real losses and apparent losses. Apply growth assumptions to the relevant customer groups, not indiscriminately to every demand component.
Represent peaks and uncertainty
Test low, central and high growth trajectories alongside plausible changes in occupancy, efficiency, climate and industrial demand. Apply peaking factors only where their definition and averaging period match your inputs. Check that a daily pattern has the intended mean and has not accidentally multiplied an already peaked demand.
Validate before extending
Back-test against a withheld period and examine seasonal and peak-period errors. For short horizons, compare machine learning forecasts with simple seasonal baselines. For long horizons, document the scenario assumptions that historical data cannot determine.
Build demand components on compatible boundaries
Suppose residential use is 1.8 ML/day and non-residential use is 0.4 ML/day. Total customer consumption is 2.2 ML/day. If real losses are defined as 12% of network input, network input is 2.2 / (1 − 0.12) = 2.5 ML/day. Adding 12% to consumption would instead give 2.464 ML/day and does not match that definition.
Keep apparent losses separate in the water balance. Under-registration changes the measured consumption estimate, but it is not the same physical escape of water as a leaking main. Avoid counting a correction both as increased customer demand and again in a residual loss allowance.
Use a common period for source meters, customer billing and storage change. Billing cycles may overlap months, and missing meters can bias a simple sum. Record the reconciliation gap rather than forcing all unaccounted volume into one cause.
Translate annual growth into spatial scenarios
Forecast population, occupancy and customer class at the scale needed for the decision. Separate committed developments from uncertain timing and ultimate capacity. Allocate growth to plausible service locations, not equally across every model node. A small growth area at a constrained elevation can matter more hydraulically than a larger addition near the source.
Distinguish customer growth from changes in per-customer use. Appliance efficiency, climate, restrictions and industry composition can offset or amplify population growth. Use scenarios that preserve plausible combinations and explain the rationale for each. A historical trend cannot independently determine a future policy or development outcome.
Apply peak factors without double counting
A volume per day and an instantaneous peak flow have different meanings.
In a teaching example, average input of 2.5 ML/day is 28.94 L/s. If a documented maximum-day factor is 1.6, maximum-day volume is 4.0 ML and its average rate is 46.30 L/s. If the maximum-hour factor within that maximum day is 1.8, the peak-hour rate is 83.33 L/s.
These factors are illustrative. Establish whether a published or measured peak-hour factor is relative to annual average, maximum-day average or another baseline before applying it. If it already refers to annual average, multiplying it by a maximum-day factor may count the same effect twice.
A pattern should also conserve the intended daily volume. For equal hourly intervals, a mean-one pattern preserves the base demand over the day. For unequal intervals, use a duration-weighted mean. Separately consider whether different customer classes peak at the same time.
Use sensitivity to choose the next investigation
Vary uncertain inputs and observe which ones change a service conclusion or investment trigger. If a trunk upgrade is needed only under a very high industrial-demand case, investigating that customer commitment may be more useful than refining household demand to another decimal place.
Record sensitivity to growth timing, loss assumptions, peak coincidence and climate. Report forecast intervals or scenario ranges with their interpretation. A range based on three planning narratives is not automatically a probabilistic confidence interval. The forecasting exercise addresses short-term chronological validation; the long-term scenario question remains an engineering and planning judgement.
Sources & further reading
- Best Practice Modelling Guidelines ↗eWater · 2011
External source · Checked 24 September 2026 - Improving short-term water demand forecasting using evolutionary algorithms ↗Stańczyk, Kajewska-Szkudlarek, Lipiński & Rychlikowski · 8 August 2022
External source · Checked 24 September 2026
Source findings are distinguished from editorial interpretation. Apply current local criteria and project evidence when making engineering decisions.