Time Series Forecasting of Natural Inflow in Hydroelectric Power Plants Using Hyper-Tuned Temporal Fusion Transformer With Hodrick–Prescott Filter

Abstract

The scheduling of the operation of the electricity system in Brazil is based on multi-criteria optimization that takes into account the forecast of the level of the dams of the hydroelectric plants, this variation is evaluated by the soil moisture active passive model. Considering the advances in using deep learning to forecast time series variations, this paper proposes a hybrid method for forecasting dam level variations. In particular, the temporal fusion transformer (TFT) is used for prediction with the Hodrick–Prescott filter for denoising. To enhance the model's performance, its hyperparameters are optimized by the Optuna framework based on the tree-structured Parzen estimator. For benchmarking, the multilayer perceptron, long short-term memory, recurrent neural network (RNN), Dilated RNN, temporal convolutional neural, neural hierarchical interpolation for time series forecasting, deep non-parametric time series forecaster, and the standard TFT are considered. The results show that the proposed model can make predictions with high performance compared to other methods, being 29.12% better than the second-best model, and 59.22% better than the original TFT model for very short-term forecasting, making it a promising alternative to be used as additional information for planning the operation of the electrical power system.