Multiple models
How we built an intelligent forecasting layer for energy sector
<p data-pasted="true">How ACTUM Digital is building an intelligent aggregation layer that learns from the strengths and weaknesses of an energy provider's existing forecasting models to produce a single, more accurate prediction of electricity consumption.</p>
At a glance
Continuous learning
the system adapts as new data arrives
15-min to monthly
accuracy tracked across all planning horizons
Built to scale
ready for multiple commodities
Why forecasting accuracy matters in energy markets
<p data-pasted="true">Unlike most commodities, electricity must be produced and consumed almost simultaneously. Entities in the electricity grid (producers, traders, distributors, and the transmission system operator) depend on forecasts of consumption, renewable production, electricity prices, and grid load to plan operations, buy and sell power, and reduce imbalances on the network. </p><p>Good forecasts translate directly into better resource utilization and lower spend on balancing (regulation) energy. The inverse is also true and considerably less forgiving: when a market participant's actual consumption imbalances from its forecast portfolio position, it is charged a financial compensation for the imbalance it caused. Forecasting accuracy, in other words, is not an abstract data science metric in this industry. It is a direct line item on the balance sheet.</p><p></p>
Challenge
Energy market participants today typically run several forecasting models in parallel, each looking at future trends from a different angle. Some models are better at capturing historical consumption patterns, others react more sensitively to short-term fluctuations or external factors like weather. This diversity of approaches is, in principle, a strength - more perspectives, more information.
In practice, it creates a different problem: multiple models, multiple answers, no single source of truth. The individual models frequently disagree with each other, and the accuracy of any single one shifts over time depending on conditions. Our analysis of the client's existing models showed that no single model was consistently the most accurate across all situations. Each one had its own strong and weak periods, and its relative reliability changed with seasonality, weather, and other external conditions.
The real challenge wasn't choosing a better model. The actual question was how to make sense of several signals whose individual value changes dynamically depending on context, and turn them into one forecast a business can actually plan around.
Why this gap was costly to leave unsolved
Even small discrepancies between planned and actual electricity consumption can lead to financial penalties, the need to purchase balancing energy, and higher operating costs overall. This is exactly where the opportunity for a new approach emerged. Not one that tries to replace existing forecasting models with a single new "best" model, but one that intelligently combines the models that already exist.
Solution
<p data-pasted="true">We are designing and validating an <strong>AI forecasting layer</strong> that sits on top of an energy provider's existing prediction models rather than replacing any of them. The system continuously compares each individual forecast against actual consumption values, tracks each model's historical accuracy, and identifies recurring error patterns. Based on this, it dynamically adjusts how much weight each model's prediction carries and produces a single consolidated forecast that is, at any given moment, the closest available estimate to reality.</p><p>Crucially, the model does not rely solely on the input forecasts themselves. External data, most importantly meteorological data, which has a substantial influence on electricity consumption behavior, also feeds into the system. The result is a model that works with a network of signals rather than a single point of view, continuously evaluating and combining them into one output.</p><p></p>
The solution included
Continuous model evaluation
<p data-pasted="true">Constantly compares each forecasting model's predictions against actual results to track who's performing well right now.</p><p></p>
Dynamic, error-aware weighting
<p data-pasted="true">Detects recurring error patterns per model and automatically adjusts its weight, no manual rules required.</p><p></p>
External data integration
<p data-pasted="true">Adds meteorological and other auxiliary data alongside the forecasts, since weather strongly drives electricity consumption.</p><p></p>
A genuinely adaptive ensemble
<p data-pasted="true">Keeps learning from new data, continuously refining model weights instead of staying fixed after training.</p><p></p>
A general-purpose architecture
<p data-pasted="true">Built generically for electricity and gas, consumption and production. Currently tested on electricity consumption.</p><p></p>
Understanding MAPE: how forecasting accuracy is measured
Forecasting accuracy in this project is measured using MAPE (Mean Absolute Percentage Error): a standard metric that expresses how far a forecast deviates from the actual value, as a percentage. The lower the MAPE, the more accurate the forecast.
MAPE is calculated across multiple time horizons: individual 15-minute intervals, full days, weeks, and months. Looking at accuracy across all of these horizons, rather than just one, matters because a model can perform very differently at a granular, short-term level than it does in daily or weekly aggregate, and energy market participants need to plan and trade across both timeframes.
Result
<p data-pasted="true">This forecasting layer is an in-house product currently in development, being validated through a pilot with a real energy sector client on real consumption data. In testing on electricity consumption forecasting, the consolidated forecast produced by the AI aggregation layer measurably <strong>improves MAPE</strong> compared to the individual forecasting models feeding into it, confirming that the ensemble approach is working as intended for the use case currently being validated.</p><p>While the current pilot focuses on electricity consumption forecasting, the underlying architecture is already designed to support multiple commodities and forecast metrics (consumption, production).</p><p>From a business perspective, a more accurate consolidated forecast translates into lower imbalance risk, more efficient electricity procurement and trading, lower balancing energy costs, and more stable operational planning. In an environment where every percentage point of forecasting error has a direct financial impact, improved forecasting accuracy is not just a data science achievement—it is a <strong>measurable business outcome</strong>.</p>
<p data-pasted="true">The underlying principle is simple, even if the system behind it is not: instead of searching for one perfect model, the system learns to understand when and why each individual model performs better or worse, and uses that understanding to build a more accurate whole. The result is an adaptive mechanism that keeps improving with every new observation and progressively reduces uncertainty in operational decision-making.</p><h3>Business Impact</h3><ul><li>Lower balancing energy costs</li><li>Reduced forecasting risk</li><li>Better purchasing decisions</li><li>Improved operational planning</li><li>Increased trust in forecasting outputs</li></ul>
Why this approach matters to energy market participants
Reduced imbalance risk
<p data-pasted="true">More accurate forecasting supports better planning and helps minimize the financial impact of consumption deviations.</p><p></p>
One forecast, multiple perspectives
<p data-pasted="true">The solution combines and validates multiple forecasting models to produce a single, more accurate forecast.</p><p></p>
Continuous learning from real-world outcomes
<p data-pasted="true">By comparing forecasts with actual consumption, the model adapts to changing conditions and improves over time.</p><p></p>
Designed for scale
<p data-pasted="true">Forecasts are delivered in a format ready for integration into reporting, analytics, and operational workflows.</p><p></p>
Next steps for your business
Curious whether an AI-driven forecasting or aggregation layer could improve decision-making in your own data-intensive operations? Contact us to discuss your use case, or explore our AI and Data services to see how we approach forecasting, optimization, and decision-support challenges across energy, utilities, and other data-rich industries. You can also browse our other case studies for more examples of how we put AI to work for our clients.