Hybrid emerging model predictive data-driven forecasting of three-phase electrical signals of photovoltaic systems using GBR-LSTM

Abstract

In numerous industrial contexts, precise analysis and forecasting of electrical signals within three-phase systems are indispensable. As a result, this work presents DeepPhase, a hybrid framework that combines Long Short-Term Memory (LSTM) neural networks with gradient-boosted regression (GBR) to predict the current, voltage, and power of electrical signals. The performance of the model is evaluated in comparison to benchmark models, namely Bidirectional LSTM (BiLSTM), K-Nearest Neighbors (KNN), and LSTM, which utilize essential Key Performance Indicators (KPIs). As demonstrated by its highest Coefficient of Determination (R2) of 0.999, Mean Absolute Error (MAE) of 6.94 × 10−5, Mean Absolute Percentage Error (MAPE) of 0.07 %, and Root Mean Square Error (RMSE) of 0.000156, DeepPhase consistently exhibits predictive precision. For Three-Phase Current, MAE is 2.13 × 10−3, MAPE is 0.01 %, RMSE is 0.062432, and R2 is 0.960596; and for Three-Phase Voltage, MAE is 9.52E-03, MAPE is 0.03 %, RMSE is 0.014, and R2 is 0.999. The results of this study highlight the effectiveness of DeepPhase in analyzing the dynamics of complex Three-Phase electrical signals. This has significant implications for improving decision-making and control strategies in complex electrical systems.