A Hybrid CNN–LSTM Model with Attention Mechanism for Photovoltaic Power Estimation

Abstract

In this study, a hybrid deep learning model is proposed for short-term active power estimation using PV (Photovoltaic) inverter current data. The model was developed to better learn time dependencies and achieve better performance in active power estimation. It combines convolutional neural networks (CNN) with bidirectional long-short-term memory (BiLSTM) architecture and is developed with Multi-Head Attention and Squeeze-and-Excitation (SE) blocks. In this study, 20 inverter current channels (PV1–PV20) and average, sum, and ratio-based features derived from these currents are used. Samples collected at a 5-minute resolution between May 2024 and July 2025 were subjected to data cleaning and rescaling processes; then, the active power of the next step was estimated based on 60 time-steps of historical data. The model was trained using the AdamW optimization algorithm and Huber loss, and performance values of MAE = 2.14 kW, RMSE = 5.30 kW, MAPE = 8.26%, and R² = 0.963 were achieved. The results demonstrate that high-accuracy short-term forecasting can be achieved using only inverter current data, without the need for weather data. The proposed model offers an effective approach for applications such as grid stability, inverter-based monitoring, energy management, and instantaneous production planning.