Прогнозування електроспоживання на базі нейронної мережі

Abstract

In this article is considered a problem of predicting discrete values of power consumption of electric energy. Since the volume of the consumption charts database must be sufficiently large to allow prediction to be reliable, it is advisable to apply modern mathematical approaches to processing these graphs, in particular, artificial neural networks. So it is solved by using one of the methods of artificial neural networks - the restricted Boltzmann machine. Restricted Boltzmann machine is a stochastic artificial neural network, the training of which is carried out without a teacher according to the algorithm of the reverse error propagation. Compared to other structures of neural networks, a restricted Boltzmann machine enables more efficient training algorithms than is possible for a general class of neural networks, in particular, the contrast divergence algorithm. This algorithm performs Gibbs sampling, which is used in the gradient descent procedure to calculate the refined weighting coefficients. In this method forecasting is carried out by adjusting the weight functions of the layers of the neural network in such a way so if optimized weighting is used, we minimize the Kulbock-Leibler distance between the distribution of input data and model approximation. In each phase of network training, weighting coefficients are calculated and modified based on their current mathematical values using the appropriate mathematical functions such as, one step of the Gibbs sample (updating the weights of all hidden and visible neurons). Accordingly, the weighed sum of the values of the output signals of the neurons, which goes on to the activation block, also changes. Depending on the function embedded in the activation block, an output signal is generated. Consequently, the result of each phase of the training is the change in the weighting factors and parameters of the activation function. As a result, after training, the system is able to effectively predict the time dependencies of the power consumption of electrical energy, which is important when developing a power management system for industrial and household objects. After studying such a system on the current database of graphs of daily electricity consumption, it will be able to make predictions within the specified error. Application of the proposed approach is effective in developing a system for managing the consumption of both large and small industrial and domestic objects, in particular Micro Grid. Ref. 5, Fig. 2