Abstract:
This study examines domestic water
demand forecasting in the Aththidiya South Grama
Niladhari Division, Colombo, Sri Lanka, where
urbanization, climate variability, and ageing
infrastructure are straining water supply systems.
Accurate demand prediction is critical for effective
planning and resource management. Traditional methods
such as linear regression and ARIMA are limited by
assumptions of linearity and stationarity, restricting their
ability to capture the complex, dynamic behaviour of
urban water consumption. To address this, the study
develops a two-stage forecasting framework combining
machine learning with classical time-series modelling.
First, a Random Forest regression model identifies and
ranks key predictors of domestic water demand using
socio-economic, climatic, and infrastructure-related
variables. Second, a Long Short-Term Memory (LSTM)
neural network forecasts monthly water demand,
capturing temporal dependencies and non-linear
consumption patterns. ARIMA(1,1,1) and
SARIMA(1,1,1)(1,1,1,7) models were developed for
benchmarking. Data were sourced from the National
Water Supply and Drainage Board, the Department of
Meteorology, and a household survey conducted in the
study area. Results show monthly household water
consumption ranging from 3–37 m3, averaging 11.3 m3.
Household size was the most influential predictor (52%
relative importance), followed by rainfall (25%) and
temperature (22%). The LSTM model outperformed
classical benchmarks, achieving an RMSE of 0.78 m3 and
MAPE of 21.32% on the test dataset. As a case study, the
framework offers promising practical value for water
resource planning in the study area, though the limited
sample and single-division scope mean that broader
generalization will require validation across additional
GN divisions and larger datasets.
