The high level of uncertainty and fluctuations, in particular in the presence of renewable energy sources, can lead to large imbalances in the real-time operation if the forecasts are inaccurate. Thereby, machine learning approaches have gained a lot of attention in the field of forecasting renewable generation and demand. To handle increasing variety and complexity of human behaviour, weather conditions and measurement data, different forecasting algorithms have been deployed in three different use cases to best reflect the specific characteristic of the location. The forecasting algorithms, namely Long-Short Term Memory (LSTM) Auto-Encoder, Hybrid LSTM-Convolutional Neural Network (CNN) and the Fuzzy-RBF-CNN have been applied to renewable generation, i.e. solar photovoltaic power and wind power, and load, whose available flexibility is critical for microgrids. The use cases include Gaidouromandra microgrid on Kythnos island in Greece, entire island of Kythnos and Bornholm island in Denmark. The prediction performances of the uncertainties are validated for all use cases, both for operational aspects and scheduling purposes for prediction horizons of 1 hour ahead, 6 hours ahead and day-ahead. For every use case, a forecasting algorithm is suggested based on the overall performance.
|27th International Conference on Electricity Distribution
|12/06/2023 → 15/06/2023
|IET Conference Publications