Study on the Impact of Sequence Length on the Medium-and Long-Term Hydro-Wind-Solar Simulation Accuracy

Aiming at the problem that the simulation effect of hydro-wind-solar systems has different adaptabilities to different models and different data lengths, a joint stochastic simulation model is first established to simulate runoff, wind power output, and photovoltaic power output data sequences of various lengths. The maximum mutual information coefficient method (MIC) is used to select the optimal time lag. Then, based on traditional data-driven models including Random Forest (RF), Support Vector Machine (SVM), Gaussian Kernel Regression (GKR), and deep learning models (LSTM, GRU, CNN, BiLSTM, iTransformer, TCN), simulation models are respectively established. The responses of the models' simulation performances to the input of historical data with different lengths are compared and analyzed, and the simulation accuracy of the models and their adaptability to hydro-wind-solar simulation objects are verified. Taking the hydro-wind-solar base in the upper Yellow River as a research case, the results show that the SVM model has relatively high accuracy, with an average R² of 0.878 in the test set. For runoff and wind power, the optimal data sequence length is 70 years; while for photovoltaic power, most models perform best at 70 years and 100 years. This study provides a reference for engineering simulation to optimize models and gives the reference values of the optimal data length under various data conditions.