ReliefF-ANN Synergistic Regression Model for Solar Microgrid Power Generation Prediction

Distributed solar energy, as a clean and renewable energy form, is playing an increasingly significant role in the power system. It can effectively integrate solar energy resources, enhance energy utilization efficiency, and improve the stability and reliability of the power system, which has become a research hotspot in the current energy field. This study proposes a regression model architecture based on ReliefF feature selection and shallow neural networks fusion for power generation prediction in distributed solar intelligent microgrid systems. Through a four-stage modeling process (data preprocessing, feature selection, neural network training, and two-stage validation), a prediction model with both physical interpretability and data-driven advantages is constructed. Firstly, using the ReliefF regression variant, four key features (temperature, atmospheric pressure, light intensity, and humidity) are selected from 8 environmental parameters, with their feature weights highly coupled with the three-layer physical mechanism of photovoltaic energy conversion (carrier dynamics layer, surface effect layer, and energy input layer); Secondly, a three-layer feedforward neural network (with 5 neurons in the hidden layer) is constructed based on the selected features, and nonlinear modeling is achieved through 1000 rounds of training, with controllable model complexity and fast convergence speed. Empirical analysis shows that the model has an RMSE of 22.72 on the test set, <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{R}^{\mathbf{2}}$</tex> of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{0. 9 0 4}$</tex>, and the prediction error is stable within 5% of the rated capacity, and the inference speed is 2.3 times faster than the full-feature model. The study finds: 1) Temperature dominates the fluctuation of power generation through bandgap modulation (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\text{Eg}=1.12 \text{eV}-\alpha \mathrm{T}^{2} /(\mathrm{T}+\beta)$</tex>), with a weight ratio of 23.8%; 2) The dynamic scheduling strategy of the microgrid leads to a decoupling phenomenon of irradiation and power generation from 10:37 to 13:21; 3) When PM2.5 < <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$75 \mu ~\mathrm{g} / \mathrm{m}^{3}$</tex>, the influence of dust deposition on transmittance can be ignored (weight <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$&lt;0.001$</tex>). This model provides an efficient and interpretable modeling approach for medium-scale renewable energy datasets and has significant application value in real-time scheduling and fault diagnosis.