Photovoltaic Generation Power Forecasting And Load Forecasting

With the increasing capacity of photovoltaic (PV) power generation systems, its impact to thegrid is becoming more and more apparent. The accurate prediction of PV power is significant toeliminate the influence to the electrical power system caused by the randomness of PV outputpower.Short term load forcasting is the promise and basic of developing the power generation andfuture planning. The accuracy of the prediction results is directly related to the economic and safeof power systems. Therefore, improving the accuracy of the photovoltaic power forecasting andload forecasting has been one of popular research fields both in domestic and overseas.The purpose of selecting similar days is to predict data in future according to historical data.The photovoltaic power data has compact connection with both load data and historical data.Consequently, the theory of similar days is applied during photovoltaic power forecasting andload prediction models in this thesis. Similar days selection principle is employing the EuclideanDistance to determine variances between feature vectors of different days based on the establishedday feature vectors.In recent years, artificial intelligence is widely introduced in photovoltaic power and loadforecasting field. Nowadays, the BP neural network is most commonly used. BP neural network isa kind of static feed-forward neural networks and it will easily fall at the local minimum pointapplied in dynamic systems. Compared to BP neural network, Elman neural network adds anundertaking layer and it can be treated as a feed-forward neural network with local memory unitand local feedback link. It is capable of excellent dynamic characteristics, fast approaching speed,and high accuracy. Thus, the Elman neural network is better than the BP neural network forforecasting the power of photovoltaic generation.The main factors which will affect photovoltaic power generation include seasons, date type,solar radiation intensity and temperature, etc. On the other hand, the load level could be effectedby the internal laws and external meteorological factors, resulting in periodicity and random. Inthis thesis, day feature vectors are established based on two impact factors, output power and loadcharacteristics, respectively, and then similar days are got. Combing the obtained similar day dataand weather forecast data, the photovoltaic output power and load forecasting models areestablished through the Elman neural network. Experimental results show certain advantages ofthe built model compared to traditional forecasting models.Solar radiation intensity is the most direct factor which will affect the generated power and theoutput power variation of PV array is almost identity to the solar radiation intensity curve. On the basis of predicted values of the solar radiation intensity, the forecasting models through the linearextrapolation method and the I/V characteristic of solar cells are established. Experimental resultsindicate that the accurate prediction of solar radiation intensity is significant to improve theaccuracy of photovoltaic power generation forecasting.The load level has obvious regularity other than PV output power and it has a closerrelationship with historical data. By selecting the similar day, the simple and practical model forultra-short-term load forecasting is established based on the linear extrapolation method. Theproposed model possesses simplicity, feasibility, and high prediction accuracy verified byexperimental results.