Hour-Ahead Load Forecasting for Flexibility Management in Energy Communities

Abstract

This study investigates the application of different Long Short-Term Memory (LSTM) architectures for hour-ahead load forecasting to support flexibility management in local energy communities. Four architectures are analyzed: Multi-layer LSTM, Unidirectional LSTM, Autoencoder LSTM, and Bidirectional LSTM. The dataset comprises hourly load measurements collected from 2018 to 2021 for a suburb in Poland, with a maximum observed load of 1.68 MW. Forecast performance is evaluated using Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and Mean Bias Error (MBE). Results show that all models provide closely comparable performance, with RMSE values ranging from 0.031 MW to 0.034 MW, corresponding to approximately 1.89% to 2.06% of the maximum load. The Unidirectional LSTM achieved the best results, exhibiting the lowest RMSE (0.031 MW) and minimal bias. Forecasts remained accurate across different seasons, effectively capturing daily load dynamics and peak periods. The findings emphasize that simpler LSTM architectures can deliver highly competitive and computationally efficient forecasting performance, making them well-suited for real-world flexibility management applications in energy communities.