Enhancing PV power forecasting through feature selection and artificial neural networks: a case study

Area of Science:

• Renewable Energy Systems
• Artificial Intelligence in Energy
• Data Science for Power Grids

Background:

• Growing demand for reliable renewable energy forecasting.
• Need for improved accuracy in photovoltaic (PV) power output prediction.
• Challenges in managing solar energy due to its intermittent nature.

Purpose of the Study:

• To enhance photovoltaic (PV) power forecasting accuracy.
• To systematically integrate feature selection techniques with artificial neural networks (ANNs).
• To identify the most relevant predictors for PV output.

Main Methods:

• Employed feature selection methods: ReliefF, minimum correlation, Chi-square test.
• Developed and tested two predictive models: Multilayer Perceptron (MLP) and Long Short-Term Memory (LSTM) networks.
• Utilized a real-world PV dataset from southern Algeria.

Main Results:

• Feature selection significantly improved forecasting accuracy for both MLP and LSTM models.
• ReliefF with MLP achieved a normalized Mean Absolute Error (nMAE) of 9.21% and R² of 0.9608.
• Chi-square selected features with LSTM resulted in an nMAE of 9.29% and R² of 0.946.

Conclusions:

• Careful feature selection enhances ANN model performance for PV forecasting.
• Feature selection reduces model complexity and improves generalization capabilities.
• Findings offer valuable insights for efficient solar energy management and grid stability.