KETEP 세계에너지시장정보

1.The invention discloses a photovoltaic power generation power prediction method based on a branch evolution neural network. Step 1: obtaining the generated power and the power influence factor time series data of the photovoltaic power generation field. Step 2: pre-processing the time series data, including performing data dimensionality reduction and training set construction. In step 3, a circular neural network model is established, and the model is trained using the training set. In step 4, real power impact factor data is obtained, and the generated power value is predicted using the recurrent neural network model. Step 5: obtaining the true value corresponding to the predicted power generation power value and continuing the incremental training of the model. The circulating neural network model is Elman neural network model and comprises four layers of structures: an input layer, an implicit layer, a bearing layer and an output layer. An Elman neural network model has a spatial expression in the non-linear state. X (t) ＝ f (w)2 xc (t)+ w1 U (t-1))xc (t) ＝ x (T-1)Y (t) ＝ g (w)3 X (t))Wherein t represents time and xc Output vector and w for receiving layer1 . w2 And w3 X (t) represents an input vector of t - type timing network and u (t) represents t-time implicit layer output vector, y (t) represents the output vector of the network output layer of the network model under the effect of the external input time sequence x (t); f (·) and g (•) represent the nonlinear vector function composed of the excitation function of the output layer unit and the implicit layer unit. In step 3, training the model is performed by using a branch particle swarm algorithm to perform parameter optimization training. Step 301: The parameters to be optimized in Elman neural network are divided into two groups according to an implicit layer neuron. At step 302, packet training is performed, including. In step 30201, two sets of parameters to be optimized are initialized to random values respectively, and the remaining to-be-optimized parameters are set as fixed values, thereby generating two random populations. In step 30202, the two populations are updated and optimized with a particle swarm algorithm, respectively. Step 30203: respectively, let two sets of training reach the preset maximum number of iterations or reach the specified precision. Step 303, performing global training, including performing global training. Step 30301: respectively selecting the optimal half of the two populations, interoperation of the optimal half of the two populations, and combining the two populations into one population. Step 30302: updating and optimizing the one population with a particle swarm algorithm. Step 30304: let the global training reach the preset maximum number of iterations or reach the specified precision. Step 304: selecting an optimal individual, calculating an optimization value for each parameter and a parameter value of the threshold value of the output layer neuron.