Application of time-space correlation correction model in ultra-short-term distributed photovoltaic power prediction

Reliable distributed photovoltaic (PV) power prediction is essential for grid managers to obtain output data and mitigate effects. An advanced model that incorporates AI and the concept of time-space correlation is introduced in this paper, addressing both the current neural network algorithms' imprecision in prediction and the insufficient investigation into the time-space correlation of stations. Firstly, a Bayesian-optimized LSTM is employed to construct a model specifically designed. This approach leverages the Bayesian optimizer to mitigate overfitting issues within the model. Secondly, based on the above prediction model, a time-space correlation correction model is proposed by resolving the intra-day multi-scale time coupling of PV power and the multiple space coupling of centralized-distributed sites. Finally, the LSTM prediction model, augmented by the time-space correlation correction model, has been demonstrated to be more accurate in predicting than other models. The MAE was enhanced by 6.38%, RMSE by 5.58%, and MAPE by 4.9% when the algorithm and model were implemented in a distributed PV power plant in Australia, as evidenced by the case study results.