Research On Power Forecasting Method For Photovoltaic Microgrid

The non-emissive feature of solar energy harvesting leads to extensive installations and penetration of photovoltaic(PV)power plants in power grids.With the high integration of PV power,the intermittent and volatile features of photovoltaic power poses significant challenges to the reliable and economic operation of power system.In order to smooth out the fluctuations in PV power and inprove the efficiency of PV generation,more traditional generation resources for rapid ramping and reserve should be scheduled.Therefore,it is particularly important to find out how to maximize PV power consumption with safe and economical operation of power grid.Accurate PV power prediction is one of the effective ways to solve this problem.Firstly,this paper analyzes the characteristics of PV power from three different time scales of atmospheric motion:long,medium and short.Based on the regular feature of irradiance caused by earth’s astronomical movement,the overall attenuation trend of weather condition and the disturbance of local cloud,PV power is decomposed into deterministic cycle variable,weather type attenuation factor and stochastic fluctuation variable in this paper.Then,the mathematical properties of each variable has been studied.In this paper,PV forecasting system consists of two parts:short-term forecasting module and ultra-short-term forecasting module.First,a short-term forecasting method for PV microgrid based on deep learning is proposed.As one kind of deep learning method,Long Short-term Memory(LSTM)network can capture the nonlinear and time-varying parameters of PV power and memorize the previous information of time series.We can find the correlation of time series and learn the inherent laws of time series.Therefore,this paper chooses LSTM network as the short-term forecasting model and outputs the predicted value directly.The simulation results show that the forecasting model based on LSTM network can learn the data features well and the short-term prediction results are pretty good.Then,this paper proposes an ultra-short-term prediction method with an improved weather classifier.The photoelectricity conversion model converts the predicted value of irradiance into the predicted value of PV power.Numerical weather prediction(NWP)is not adopted here,because the time resolution of NWP is not high enough for ultra-short-term forecasting.Besides,it is not suitable for PV microgrid because the generally installed capacity of PV microgrid is small and using NWP will greatly increase the cost.Therefore,this paper proposes the weather pattern recognition method based on historical data.A weather type classifier based on an integrated classification algorithm(Adaboost)was designed to classify typical weather types at low cost.At the same time,the attenuation coefficients of typical weather types were introduced to refine the attenuation state under cloudy and rainy conditions.Then,a multi-order weighted Markov chain ultra-short-term prediction model based on error sequences was established.Finally,a practical photoelectric conversion model was used to convert the prediction value of irradiance into prediction value of PV power.Simulation results show that the integrated weather classifier proposed in this paper can recognize the feature of original PV power data better and its classification accuracy is higher than simple base learner.Besides,the improved ultra-short-term forecasting model proposed in this paper improves the prediction accuracy under non-sunny condition.