BCAM – Probabilistic Predictions for Energy Management Based on Continuous Learning
BCAM
Sector: Energy
Business Case
Obtaining accurate energy demand predictions is difficult due to the continuous changes in consumption patterns (concept drift). Accurate forecasts can optimise the planning of energy production as well as its distribution, resulting in efficient generation and minimising the risk of supply interruptions. This project develops machine learning algorithms that obtain accurate predictions by effectively adapting to temporal changes. Probabilistic predictions quantify the uncertainty of predictions, which makes them an essential requirement for making optimal decisions.
Objectives
Develop energy prediction methods with theoretical guarantees that obtain probabilistic predictions and are based on continuous learning. Probabilistic forecasts quantify the uncertainty in demand and continuous learning allows for adapting to the concept drift. The methods developed use hundreds of models that are continuously learning (continual learning) and obtain predictions over time, always using the most recent data.
Use case
Flexible prediction methods in temporal granularity of data, prediction time horizon and external variables. In addition, the methods developed are based on minimax classification that allow for predicting increases and decreases in the price of energy. The techniques developed obtain predictions using models updated with the most recent real data and have theoretical performance guarantees.
Infrastructure
On Premise
Technology
Machine learning and deep learning
Data
Multiple public time series datasets corresponding to regions with different sizes and consumption patterns that change over time. The datasets include hourly energy consumption as well as external variables that affect consumption, such as temperature.
Resources
Multiple public time series datasets corresponding to regions with different sizes and consumption patterns that change over time. The datasets include hourly energy consumption as well as external variables that affect consumption, such as temperature.
Difficulties and learning
Challenges associated with energy forecasting models include the availability of quality public historical data, the number of external variables affecting energy consumption, uncertainty in demand, and concept drift. Learning to develop energy prediction models involves understanding the energy domain, acquiring data management skills, considering relevant external factors, and developing models that capture uncertainty and adapt to concept drift.
KPIs (business impact and metrics of the model)
Predicciones precisas de la demanda de energía permiten optimizar la gestión de recursos, mejorar la eficiencia operativa y maximizar los beneficios económicos. La precisión de las predicciones obtenidas es evaluada utilizando métricas como el error percentual medio absoluto (MAPE) y la raíz cuadrada del error cuadrático medio (RMSE). Además, evaluamos la calidad de las predicciones probabilísticas utilizando métricas como el pinball loss que penaliza las discrepancias entre las predicciones y los valores reales en diferentes cuantiles.
Funding
IA4TES project, Ministry of Economy and Digital Transformation. Next Generation EU. “Artificial intelligence for energy management” project, IBERDROLA Foundation Chargers+ and Twin-net projects, ELKARTEK Program, Basque Government
Collaborators, Partners
IBERDROLA (foundation, R&D&I department, finance department) TECNALIA Research and Innovation
