Electricity Demand Forecasting with Machine Learning : A Case Study on the Implications of Data Availability and Economic Impact for German Municipal Utilities

Sammanfattning

Background. Artificial intelligence (AI), and more specifically machine learning (ML), is increasingly gaining importance for its potential to improve forecasting, operational efficiency, and decision-making in the energy sector. While several use cases have been identified, adoption remains limited due to barriers such as poor data availability, a lack of AI expertise, and regulatory uncertainty. German municipal utilities, in particular, face challenges stemming from limited access to high-resolution consumption data and slow smart meter rollout. Beyond technical accuracy, deviations between predicted and actual electricity consumption can lead to significant financial impact, affecting procurement and costs under Germany’s Excess and Shortfall Quantity Settlement (ESQS) mechanism. Yet, existing research has largely focused on technical aspects, leaving a gap in understanding the economic implications of ML-driven forecasting. Aim. This thesis investigated the applicability and economic implications of ML-based electricity demand forecasting in a German municipal utility, focusing on (1) the extent to which ML models can support forecasting under current data constraints, and (2) how forecast accuracy influences economic performance. Method. A case study approach was applied using an adapted CRISP-ML framework, integrating both technical and economic evaluation. A structured experimental design assessed data readiness across three dimensions: (i) informational content, (ii) volume, and (iii) granularity. Forecasting models were evaluated using six-fold year-wise cross-validation (2018-2023), subsampling (20 %, 60 %, 100 %), and varying feature richness. Evaluation criteria included predictive accuracy (sMAPE, MAE), consistency (min-max range, standard deviation), and computational efficiency. Economic implications were simulated based on procurement and ESQS outcomes aligned with actual procurement and settlement logic. Results. While ML models such as Random Forest, LightGBM, and Deep Neural Networks were feasible and stable, they did not consistently outperform the utility’s existing method. The primary limitation was low data granularity, with billing-period-level consumption and aggregated external features hindering the learning of seasonal or behavioral patterns. Increasing data volume had a negligible effect on accuracy but increased computational demand significantly. Despite similar sMAPE values, models showed varying economic outcomes due to consistent negative prediction bias in some models. In the prevailing market context, underforecasting led to reduced procurement volumes but higher balancing costs. Notably, the utility’s existing method achieved the most favorable average net cost, outperforming even idealized prediction scenarios in certain years. Conclusion. This study highlights the importance of aligning technical forecasting methods with both data readiness and economic context. The economic simulations revealed that forecast accuracy alone does not determine financial outcomes, model biases, market mechanisms, and temporal dynamics play a critical role. Overall, the findings suggest that machine learning can be a valuable tool in municipal electricity forecasting, but its effective use requires enhanced data infrastructure, strategic integration, and context-specific model evaluation.