A federated learning model with the whale optimization algorithm for renewable energy prediction

• FedWOA: a federated learning model for predicting prosumers renewable energy production. • Whale optimization algorithm for global model inference from the weights of local LSTM models. • K-Means clustering algorithm to address the problem of non-IID data in renewable energy production. • Better energy prediction accuracy compared to FedAVG with good convergence and reduced loss. Federated prediction models for energy prosumers create a global model by combining insights from local machine learning models trained on-site without centralizing the data. For time series energy data, this collaborative approach faces challenges due to the non-IID nature of the data, variations in generation patterns, the high number of model parameters, and convergence issues, leading to poor prediction accuracy. This paper introduces a novel federated learning model, FedWOA, which uses the whale optimization algorithm to determine optimal aggregation coefficients based on the local model weight vectors by pondering the updates considering the model performance and data dimensionality construct the global shared model. To handle the non-IID data the prosumers were clustered based on the similarity of their energy profiles using K-Means. FedWOA improves the prediction quality at the prosumer site, with a 16 % average reduction of the mean absolute error compared to FedAVG while demonstrating good convergence and reduced loss.