Application of GRNN neural network in power system load forecasting

Electricity load forecasting is a prediction of future electricity demand based on known electricity demand, taking into account political, economic, climate and other related factors. Load forecasting includes two meanings: prediction of future demand (power) and prediction of future electricity consumption (energy). It provides a basis for power system planning and operation, and is an important part of power system planning and dispatching; at the same time, it determines the annual power supply and consumption of each power supply area, the maximum load of power supply and consumption, and the total load development level of the planning area, and determines the load composition of each planned year. The current forecasting methods include trend analysis, regression analysis, exponential smoothing, unit consumption method, gray model method, load density method and elasticity coefficient method. The load curve is a nonlinear function related to many factors. Neural network is a suitable method for extracting and approximating this nonlinear function. The advantage of neural network is that it has the ability to simulate multiple variables without making complex related assumptions on the input variables. It does not rely on expert experience, but only on observed data; it can extract and approximate the implicit input/output nonlinear relationship through learning during the training process. Studies have shown that using neural network technology for short-term load forecasting of power systems can achieve higher accuracy.

1 Data Source
In order to make better use of electric energy, it is necessary to do a good job in short-term forecasting of power load. Here, the active load values ​​of a power-deficient city from July 10 to July 20, 2010, and the meteorological characteristic state quantities from July 11 to July 21, 2010 are used as network training samples to predict the power load on July 21. The data are shown in Table 1, and all data have been normalized.

In the sample, the input vector is the actual power load data on the forecast day, and the target vector is the power load on the forecast day. Since these data are all actual measured values, the network can be effectively trained. If we start from the perspective of improving network accuracy, we can increase the number of network training samples on the one hand, and increase the input dimension on the other hand. At present, there is no universal method to determine the number of training samples. It is generally believed that too few samples may make the network's expression insufficient, resulting in insufficient extrapolation ability of the network. Too many samples will cause sample redundancy
, which will increase the burden of network training and may cause the network to overfit due to excess information. Therefore, the sample selection process needs to pay attention to representativeness, balance and the characteristics of the power load itself, and reasonably select training samples.

[page]2 Network creation and training
2.1 Establishment of GRNN neural network model
GRNN neural network has been applied in system identification and predictive control. The structure of GRNN two networks is shown in Figure 1.

The first layer is the input layer, and the number of neurons is equal to the number of input parameters. The second layer is the radial basis function hidden layer, and the number of neurons is equal to the number of training samples. R represents the dimension of the network input, and Q represents the number of neurons in each layer of the network, and also represents the number of training samples. The transfer function of the hidden layer is the radial basis function, and the Gaussian function is usually used as the transfer function. The transfer function includes a smoothing factor. The smaller the smoothing factor, the stronger the sample approximation ability of the function. Conversely, the smoother the basis function. The third layer is a simple linear output layer.
This paper mainly studies the use of the GRNN neural network in the artificial neural network toolbox in the MATLAB environment to predict the power load. Since the establishment and prediction of the GRNN network are carried out simultaneously, there is no need to train the network specifically. The parameters required for the establishment of the network are the training sample input data and the training target data. Since the smoothing factor affects the network performance, the GRNN network is to find the optimal smoothing factor, starting from 0.05 and increasing by 0.05 each time to determine the optimal value.
2.2 Establishment of BP network prediction model
BP (Back Propagation) network was proposed by a team of scientists led by Rumelhart and McCelland in 1986. It is a multi-layer feedforward network trained by the error back propagation algorithm and is one of the most widely used neural network models. BP network can learn and store a large number of input-output mode mapping relationships without revealing the mathematical equations describing this mapping relationship in advance. Its learning rule is to use the steepest descent method to continuously adjust the weights and closed values ​​of the network through back propagation to minimize the sum of square errors of the network. The topological structure of the BP neural network model includes an input layer (imput), a hidden layer (hide layer) and an output layer (output layer). Its network structure is shown in Figure 2.
A three-layer BP network is selected. On the day before the forecast, the power load is measured every 2 hours, and a total of 12 sets of load data are measured in one day. Since there will be no sudden changes between adjacent points of the load curve, the value of the next moment must be related to the value of the previous moment, unless there are special circumstances, so the real-time load data of one day is used as the sample data of the network.
Since the power load is also related to environmental factors, such as the highest and lowest temperatures, etc. Therefore, it is also necessary to obtain the highest temperature, lowest temperature and weather characteristic values ​​(sunny, cloudy or rainy) on the forecast day through weather forecast and other means. Use this form to represent the weather characteristic value: 0 represents sunny, 0.5 represents cloudy, and 1 represents rainy. Here, the meteorological characteristic data on the day of the power load forecast is used as the network input variable, so the input variable is a 15-dimensional vector. The target phasor is the 12 sets of load values ​​on the day of the forecast. That is, the load value at each hour of the day. In this way, the output variable is a 12-dimensional vector. The
input and output variables are normalized and the data is processed into data between the interval [0, 1]. The normalized data uses the following formula: According to the analysis of the factors affecting power consumption, the real-time load data of one day is taken as the network and the meteorological characteristic data on the day of the power load forecast is taken as the influencing factor. The 12 sets of load values ​​on the day of the forecast are used as the network output. Thus, the BP network is constructed.

3 Experimental results
The prediction error curve is shown in Figure 3. It can be seen from the figure that the error between the network prediction value and the true value is very small. In the BP network prediction, except for a relatively large error in the 8th time, the rest of the errors are around 0. However, compared with the GRNN network, the error value of the GRNN network is smaller.

GRNN neural network has stronger advantages than BP network in terms of approximation ability, classification ability and learning speed. In addition, GRNN network has fewer parameters to be adjusted manually, only one threshold, and the distribution density of radial basis function SPREAD can have an important impact on GRNN performance. The learning of the network depends entirely on data samples, so that the network can avoid the influence of subjective assumptions on the prediction results to the greatest extent.

4 Conclusions
This study uses GRNN neural network and BP neural network to establish power load models and predict power load. From the prediction effect, the prediction error of BP network is relatively large. GRNN neural network is effective in power load prediction. Moreover, in terms of specific network training, compared with BP neural network, since there are fewer parameters to be adjusted and only one smoothing factor, it can find a suitable prediction network faster, which has a greater computational advantage.