| With the development of new energy,the installed capacity of photovoltaic in China has grown rapidly.At present,it has become the most PV power generation and installed capacity in the world.Photovoltaic power generation has the characteristics of time variability,volatility and randomness.Even on the monthly time scale,the power generation is difficult to be controlled precisely as the traditional thermal power.Under the current power market system in China,the monthly power transaction plan of thermal power needs to be formulated after the distribution of wind power and photovoltaic.With the continuous improvement of the installed capacity of photovoltaic power in China,the error of monthly photovoltaic power generation forecast is further increased by approximate estimation.Under the condition of ensuring the abandonment rate,the implementation of the original thermal power plan has been seriously affected.Therefore,in order to ensure the safety of higher proportion of photovoltaic grid connection in the future,more accurate monthly photovoltaic power forecast is needed.Based on the above problems,this paper proposes a monthly PV generation combination prediction method based on DBN-GRU.The main research results can be summarized as follows:(1)The influence of meteorological parameters on photovoltaic power generation is analyzed by using historical meteorological data and power generation data,which provides the basis for input selection for the establishment of the subsequent prediction model.Through the analysis of the characteristics of long-term scale photovoltaic power generation,the chaos characteristics and periodic rules of photovoltaic power generation are summarized,which provides theoretical basis for the subsequent data expansion.(2)In view of the problem that the historical data of monthly photovoltaic forecast has deviation and the quantity cannot meet the algorithm demand,a method of extending the monthly PV historical data is proposed.The data correction method is used to solve the problem of inaccurate data caused by the shutdown and expansion of PV power station in a long time scale.The unknown historical power generation capacity before the construction is supplemented by the historical meteorological data before the construction.Through the analysis of the characteristics and periodicity of photovoltaic power generation,two kinds of data expansion methods are used to expand the historical data again.The photovoltaic data extension method provides data support for time series and neural network algorithm,improves the accuracy of prediction to a certain extent,and provides more stable digital characteristics for interval prediction.(3)In view of the problems of large error and low stability in monthly prediction of existing point prediction model,the deep confidence network,gate control cycle unit and ARIMA time series method are selected as the core point prediction algorithm.The deep confidence network can effectively peel off the deep characteristics of photovoltaic prediction,and the gate control cycle unit can consider the sequence continuity on the basis of ensuring the real-time input.Arima algorithm can maintain certain prediction accuracy in the case of poor accuracy of input in the second half of monthly forecast.Each of the three points forecasting methods has advantages in monthly PV forecasting,which is complementary to each other,and effectively improves the accuracy of monthly PV point prediction.(4)The combined curve is weighted by the optimal algorithm,and the initial value of the combined curve is obtained by analyzing the meteorological forecast score.The parameters of the combined curve are optimized by historical data.The simulation results show that the error of the combined prediction is 4.28%lower than that of BP neural network model,which fully verifies the feasibility and effectiveness of the prediction model.On this basis,the interval prediction of the calculation example is carried out by using the digital characteristics obtained from the data extension set,which further enriches the prediction results and facilitates the implementation of the actual project. |
