Advancing photovoltaic power forecast accuracy: A hybrid particle filter and recurrent neural network model for data-scarce scenarios toward thermal adaptation

Accurate PV forecasting is vital for grid stability, efficient energy management, and maximizing the use of solar power, especially as solar adoption grows and weather conditions remain unpredictable. This study examines particle filter-enhanced neural networks for photovoltaic power forecasting toward thermal adaptation. This study introduces a novel hybrid modeling framework that, for the first time, integrates a Sequential Monte Carlo method (specifically, a particle filter) directly into the end-to-end training process of Recurrent Neural Networks (LSTM and GRU) for solar power forecasting. In a departure from conventional post-processing applications, the neural network was treated as the state transition function within a Bayesian filtering framework. The model was trained on a distribution of state hypotheses, known as particles, which were sampled from the feature space. Through this process, the network was compelled to learn a robust mapping that is inherently resilient to the high stochasticity and noise characteristic of meteorological data. The approach enables robust forecasting performance even with limited and noisy datasets. In this study four neural network models: PF-LSTM, PF-GRU, LSTM, and GRU were tested. The models used weather data including solar irradiance, temperature, humidity, wind speed, and atmospheric pressure. Present study evaluated forecasting performance at 15, 30, and 60-min intervals. The study's key innovation is using particle filters during training. This technique samples from feature distributions to better handle weather uncertainty and improve model strength. Particle filter-enhanced models showed excellent results. PF-LSTM performed best with RMSE values of 4.07, 4.82, and 5.84 W/m 2 for the three time horizons. It maintained correlation coefficients above 0.989. Particle filter enhancement reduced prediction errors by 60–70 % compared to standard methods. PF-LSTM showed minimal bias and stable error patterns across all forecasting periods. The particle filter effectively handled weather variability. By sampling from feature distributions during training, the particle filter effectively addresses the high variability in weather data that traditional neural networks struggle to model, leading to a marked reduction in prediction errors.