A generation method of meteorological year for photovoltaic buildings based on AR-LSTM with an improved diffuse estimation model

In recent years, China's renewable energy has developed rapidly, and photovoltaic buildings are an important form of renewable energy utilization. In order to conduct better preliminary energy-saving design and assessment for photovoltaic buildings, it is necessary to establish the meteorological year for photovoltaic buildings. Firstly, regarding the research on the radiation meteorological years of existing photovoltaic buildings, most of the studies focus on the global solar radiation as a meteorological parameter, while ignoring the two components of the global solar radiation when it reaches the horizontal surface: direct radiation and diffuse radiation. These two components have significant impacts on photovoltaic power generation under different weather conditions. Although there are a large number of existing studies on diffuse radiation estimation models, the current research does not take into account the influence of various factors in the atmosphere and the transparency of clouds. The accuracy of the diffuse radiation estimation model still needs to be improved. Second, the meteorological parameters involved in constructing the meteorological year for photovoltaic buildings all have strong temporal characteristics. In particular, direct radiation and diffuse radiation have the characteristics of rapid changes and high volatility. The existing methods for generating the radiation meteorological years are unable to meet the requirements of accurately capturing the temporal characteristics, rapid changes, and high volatility of these meteorological parameters. Therefore, this paper has accomplished the following works: (1)Based on the model of the diffusion fraction and diffusion coefficient after adding precipitation. By introducing cloud transmittance and atmospheric quality coefficient as input parameters, the accuracy of the existing diffuse radiation estimation model was further improved. (2)The method combining the autoregressive model and the long short-term memory network model (AR-LSTM) is adopted. The AR-LSTM model uses AR model to capture the short-term variation patterns in the radiation data, and LSTM model to capture the long-term patterns of solar radiation changes. Selecting the typical months from the long-term data that can represent the long-term average characteristics of meteorological factors such as direct radiation and diffuse radiation, thereby constructing the meteorological year for photovoltaic buildings. This paper conducts verification experiments using four cities as examples in the solar radiation zoning: Golmud and Lhasa in Zone Ⅰ, and Kashgar and Urumqi in Zone Ⅱ. The results show:①The R2 (coefficient of determination) values of the improved diffuse radiation estimation models for the four cities ranged 0.96–0.99, showing a good fit with the actual values; the values of RMSE (root mean square error) and MABE (mean absolute deviation error) were 0.007–0.021, indicating good prediction accuracy; the value of MBE (mean deviation error) was close to the actual values for all four cities. ②The fitting degree R2 of the various main meteorological parameters (dry bulb temperature, relative humidity, direct radiation and diffuse radiation) of the meteorological year for photovoltaic buildings in the four cities to the long-term average values ranged 0.78–0.99, indicating a good fit to the long-term average values. The AR-LSTM model shows better performance in terms of direct and diffuse radiation compared to Bi-LSTM, LSTM, RF, and XGBoost model. For other meteorological parameters, the relative error range of the AR-LSTM model from the long-term average is 0.02%-4.94%. The AR-LSTM model demonstrates a stronger advantage in maintaining the long-term statistical characteristics of the key radiation input variables for photovoltaic buildings.