Deep reinforcement learning-based time series forecasting in electric energy harvesting

• Autor(a) principal: Fonte, Tiago Mateus Silveira Lourenço da
• Data de Publicação: 2023
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositórios Científicos de Acesso Aberto de Portugal (RCAAP)
• Texto Completo: http://hdl.handle.net/10773/40918
• topic: Energy harvesting; Time series forecasting; Reinforcement learning; Deep learning; Electromagnetic generator

Clean and renewable energy production methods have been highly researched over the last decades to fight climate changes caused by greenhouse gas emissions. Energy harvesters surged as a promising technique to overcome the inherent limitations of small and large scale forms of renewable energy production. However, power density, efficiency and energy management of such electric generators are highly dependent on forecasting methods that can accurately predict future energy values for varying mechanical excitations. Many forecasting methods have been researched in the past, but those based on deep reinforcement learning (DRL) have not yet been the subject of much research as they are relatively recent. In this study, a deep reinforcement learning-based time series forecasting model is developed and proven to predict future power outputs of an electromagnetic energy generator . Firstly, datasets related to the instantaneous power output of the generator were gathered and processed for training and testing procedures. Secondly, the proposed model was developed based on the deep deterministic policy gradient (DDPG) algorithm and optimized using the black hole (BH) heuristic optimization algorithm. Finally, a deep learning (DL) model, based on the long short-term memory (LSTM) algorithm was developed and optimized with the BH algorithm to establish an accuracy and computation time-wise comparison between DRL and DL. Experiments showed that the proposed model and the comparison model were equally able to accurately predict the future power values of the energy generator. Considering the chosen dataset for the optimization procedure, the proposed model was able to outperform the comparison model at the cost of training time. The mean squared error (MSE), the root mean squared error (RMSE) and the coefficient of determination (R2) evaluation measures of the proposed model were, respectively, 30.19%, 65.77% and 0.22% better than the comparison model. Moreover, the proposed model was 94.80% faster than the comparison model at predicting future power values. The findings within this investigation might achieve significant impact in both small and large-scale domains, as the validation of this forecasting model is a revolutionary gateway to the creation of inteligent and highly predictive energy generators.