PV Power Prediction Based on LSTM With Adaptive Hyperparameter Adjustment

Abstract

The randomness, volatility, and intermittence of solar power generation make it difficult to achieve the desired accuracy of PV output-power prediction. Therefore, the time learning weight (TLW) proposed in this paper is used to improve the time correlation of the LSTM network. The Fusion Activation Function (FAF) is used to resolve gradient disappearance. Learning Factor Adaptation (LFA) and Momentum Resistance Weight Estimation (MRWE) are used to accelerate weight convergence and improve global search capabilities. Finally, this paper synthesizes the improvement and proposes the AHPA-LSTM model to stabilize the convergence domain. Using actual data verification, the δMAPE indicator of the improved model is only 2.85% on a sunny day, 5.92% on a cloudy day, 7.71% on a rainy day, and only 5.8% on average. Therefore, the AHPA-LSTM model under full climate and climatic conditions has a good predictive effect which is generally applicable to the prediction of ultra-short-term PV power generation.

Keywords

• Photovoltaic output power
• ultra-short-term prediction
• long short term memory (LSTM)
• time weight decoupling
• adaptive hyperparameter adjustment