Short-Term Active Power Forecasting in a Photovoltaic Plant Using TiDE and LSTM Models

This work presents a deep learning approach for short-term forecasting of active power in photovoltaic (PV) plants operating within islanded microgrids. Three forecasting schemes were implemented in Python using the Darts library: a baseline Long Short-Term Memory (LSTM) network, a Time-series Dense Encoder (TiDE) model without covariates, and a TiDE variant with exogenous inputs (solar irradiance, ambient temperature, and reactive power). Models were trained on hourly data from a 952 kW PV system on Isabela Island (Ecuador) using 2022 records. The forecasting horizon was 28 days, produced as two consecutive 336 -hour windows (14 days each). Performance was assessed with Mean Absolute Error (MAE) and Mean Squared Error (MSE). The TiDE model without covariates matched the LSTM baseline in MAE while reducing MSE by 19.4%. Including covariates did not improve accuracy but substantially decreased the number of training epochs, indicating faster convergence. These findings demonstrate that TiDE achieves competitive accuracy with lower computational cost, thereby supporting more efficient dispatch strategies in islanded and resource-constrained microgrids.