Hypertuned wavelet convolutional neural network with long short-term memory for time series forecasting in hydroelectric power plants

Energy planning in Brazil is based on assessing the availability of hydrological resources in the future, thus guaranteeing the supply of energy based on hydroelectric generation. Currently, the definition of energy dispatch by the Brazilian National System Operator is based on the rainfall forecast calculated by the soil moisture accounting procedure (SMAP). Considering that the dam's level is associated with rainfall, a forecast based on the recorded history regarding its variation trend might bean alternative for better management of the electrical power system. In this context, machine learning-based prediction models are an alternative to enhance time series forecasting. This paper proposes a hypertuned wavelet convolutional neural network (CNN) approach with a long short-term memory (LSTM) network for precipitation and natural inflow forecasting considering the variation trend over time. The proposed hypertuned wavelet-CNN-LSTM model, considers wavelet for signal denoising, CNN for feature extraction, LSTM for time series forecasting, and hypertuning to achieve an optimized structure. With a root mean squared error of 872.1, mean absolute error of 626.2, mean absolute percentage error (MAPE) of 0.071, and symmetric MAPE of 6.78, the proposed forecasting method outperforms the currently used SMAP model and other state-of-the-art deep learning approaches.