Short-Term Photovoltaic Power Forecasting Based on Historical Information and Deep Learning Methods

Area of Science:

• Renewable Energy Systems
• Artificial Intelligence in Power Engineering
• Time Series Forecasting

Background:

• Accurate photovoltaic (PV) power prediction is vital for power system planning and intelligent grid development.
• The inherent intermittency and instability of PV power data pose significant challenges to reliable forecasting.
• Short-term PV power forecasting is essential for grid stability and efficient energy management.

Purpose of the Study:

• To develop a robust deep learning framework for accurate short-term PV power prediction.
• To introduce a hybrid model integrating Singular Spectrum Analysis (SSA), Bidirectional Long Short-Term Memory (BiLSTM) networks, and Bayesian Optimization (BO).
• To evaluate the model's performance for 7.5 min-ahead and 15 min-ahead PV power forecasting.

Main Methods:

• Singular Spectrum Analysis (SSA) was employed to decompose PV power series into manageable sub-signals.
• Bayesian Optimization (BO) algorithm was utilized for automated hyperparameter tuning of the deep neural network architecture.
• Parallel BiLSTM networks were implemented to predict individual sub-signal components, with final predictions obtained by summing the results.

Main Results:

• The proposed hybrid SSA-BiLSTM-BO model demonstrated significant error reduction in PV power forecasting.
• 7.5 min-ahead predictions achieved up to 380.51% error reduction compared to baseline methods.
• 15 min-ahead predictions showed error reductions of up to 296.01%.

Conclusions:

• The hybrid SSA-BiLSTM-BO model offers superior performance for short-term PV power forecasting.
• The framework effectively addresses the challenges posed by PV power data intermittency and instability.
• The proposed method provides a reliable tool for enhancing the stability and efficiency of intelligent power grids.