Solar irradiance prediction with variable time lengths and multi-parameters in full climate conditions based on photovoltaic greenhouse.

[Display omitted] • A method for predicting the irradiance of PV greenhouse in many cases was proposed. • Four models were discussed at variable time, data size and climatic conditions. • The instantaneous irradiance prediction under 9 weather conditions was provided. • Transformer model had the best prediction effect for dataset at one year of usage. • Pyraformer model had the best prediction effect on winter and summer. Photovoltaic power generation can provide energy for greenhouses and achieve high quality and high yield of crops. In reality, solar irradiance is fluctuating and intermittent. Thus, the key to ensure efficient photovoltaic power use under greenhouse environmental conditions is to provide an accurate prediction of solar irradiance. Yet, currently studies on irradiance prediction in terms of the variable time lengths, multi-parameter and full climate conditions are rather limited. To improve the comprehensive performance of the prediction model, this paper proposes models of irradiance time series prediction such as Pyraformer, Informer, Transformer and TimesNet. Those models were tested based on the synergistic combination of weather conditions (WC), sunshine time accumulation (STA/h), instantaneous total irradiance (ITI/(W/m2)) and irradiance daily accumulation (IDA/(MJ/m2)). The model performance was rigorously evaluated with 9 prediction lengths, 5 training days, 4 seasons and 5 days for 9 weather conditions. The results showed that the Transformer model had the best overall prediction performance for STA, ITI, IDA and WC at different time steps of 10 min to 24 h. All models were suitable for predicting time series within 1 h. However, TimesNet model was not suitable for predicting time series with steps outside 1 h. On the other hand, by using the sun combination dataset, the Transformer model had the best performance at a time step of 10 min. The mean absolute error (MAE), mean-square error (MSE), root mean squared error (RMSE) and coefficient of determination (R2) of ITI were 0.118 W/m2, 0.059 W/m2, 0.243 W/m2 and 93.9 %, respectively. When exploring the minimum dataset, with the increase of data samples, the prediction effect of TimesNet showed an increasing trend. While, Transformer had the best prediction effect for the dataset with one year of use. When exploring seasonality, Pyraformer model had the best prediction effect on winter and summer, and TimesNet had the best prediction effect on autumn and spring. Local prediction of 9 climate conditions showed that the effects of snow and dust storm were not ideal. The research results showed that the characterization factors that were closely linked to irradiance. The prediction scheme proposed in this paper combined the advantages of different time steps, different factor combination datasets, different data volumes and seasonality, which greatly improves the generalization ability of the model. This study can provide a reference for irradiance prediction and more refined management of photovoltaic (PV) greenhouse. [ABSTRACT FROM AUTHOR]