A Miningful case study in utilities/water/gas contexts
At the end of each year, the total consumed volume is a key input for operational and business estimates, including the assessment of losses and other performance indicators. However, waiting until 31 December to obtain a reliable figure limits the ability to intervene during the year. The objective of this project was therefore to estimate, in advance and with adequate precision, the total volume consumed at year-end using the historical series of meter readings available for each plant.
Miningful developed a predictive approach designed to transform reading data into reliable forecasts of annual consumption. The goal was not only to predict the final 31/12 volume, but to do so early enough to support monitoring, decision-making, and preventive actions. The target level of accuracy was ambitious: the model was expected to achieve an error below 2%, and to do so earlier than existing estimation approaches. The project required addressing the transformation from meter readings to consumptions, including breaks around 31/12 and other discontinuities.
The project demonstrated highly accurate forecasts from readings. This creates the basis for a practical forecasting workflow that can be updated during the year and integrated into a web application with real-time visualizations and warning signals.
Predictive Approach and Implementation
Miningful approached the problem as a feature-engineering and model-selection challenge. The team built a modelling pipeline capable of extracting information from historical readings and testing different predictive configurations. The solution was based on the calculation of 100+ predictive features from the reading series. These features were then selected and combined to train predictive models.
A key element of the methodology was the development of models on a monthly basis. This allowed the team to assess how early in the year reliable predictions could be obtained, and which months contained the most predictive information. The implementation evolved over three main phases.
- The work focused on monthly models, the comparison between fixed and mobile features, and feature augmentation.
- The team introduced a “delay” parameter and worked on daily consumption estimates.
- The focus shifted to feature selection, model selection, optimization across the 12 months, and harmonization of the final modelling strategy.
This staged approach made it possible to progressively refine the predictive pipeline, moving from model exploration to a more structured monthly prediction framework. The result was a system that can estimate year-end consumption at different moments during the year, while also identifying the months in which the forecast becomes most reliable.
Results and Impact
The results show that the prediction of year-end consumption from meter readings is feasible with a high level of accuracy. The broader analysis also showed that predictions reached always an error below 1% within the first months of the year.
From a business perspective, the value of the model lies in earlier visibility. Instead of waiting until year-end, the organization can obtain predictive estimates during the year and use them to support operational monitoring. This creates opportunities to identify anomalies, investigate potential losses, and take preventive action before the final annual balance is known.
