How Does Neural Network Model Capacity Affect Photovoltaic Power Prediction? A Study Case.

Area of Science:

• Renewable Energy Systems
• Artificial Intelligence in Energy
• Electrical Engineering

Background:

• Accurate forecasting of photovoltaic (PV) energy production is crucial for integrating renewable sources into existing power grids.
• Long Short-Term Memory (LSTM) networks are commonly employed for PV forecasting but suffer from high computational complexity and slow performance.
• Alternative neural network architectures are needed to address the limitations of LSTMs in short-term PV energy prediction.

Purpose of the Study:

• To evaluate and compare the performance of Multilayer Perceptron (MLP), Recurrent Neural Network (RNN), and LSTM models for short-term PV energy forecasting.
• To determine the most efficient neural network for predicting 5-minute ahead PV energy production.
• To assess the impact of computational complexity on the suitability of different models for PV energy forecasting.

Main Methods:

• Utilized historical PV system data (power, irradiation, cell temperature) from 2019-2022 in Maceió, Brazil.
• Employed MLP, RNN, and LSTM neural network architectures for 5-minute ahead PV energy production forecasting.
• Applied Bayesian hyperparameter optimization to tune each model for optimal performance comparison.

Main Results:

• MLP models demonstrated satisfactory performance in forecasting PV energy production.
• MLPs required significantly less time for both training and forecasting compared to RNNs and LSTMs.
• The performance of MLP suggests its viability for very short-term PV energy prediction tasks.

Conclusions:

• Multilayer Perceptron (MLP) neural networks present a computationally efficient and effective alternative to LSTMs for short-term PV energy forecasting.
• MLPs are a practical choice for PV energy prediction systems with limited computational resources.
• The study highlights the trade-offs between model complexity and performance in renewable energy forecasting applications.