| 摘要 |
FIELD: physics.SUBSTANCE: method is realised by a mathematical model of a neural network, which is a hybrid network with cascade connection of the Kohonen distribution layer and a predicting two-layer perceptron network; the input vector of the mathematical model of the neural network includes daily values of the gradient of local temperature field expressed by coordinates, and corresponding values of the water level over the past eight days at the prediction point; before use, the mathematical model is trained on daily 20-year data, as a result of which the Kohonen layer accumulates information on classes of the course of the values of the gradient of the local temperature field and the level of water at the prediction point by selecting clusters of values corresponding to the observed weather patterns; during prediction, an input vector is transmitted to the input of the Kohonen layer, the input vector including daily values of the gradient of the local temperature field and corresponding values of the water level over the past eight days; at the output of the Kohonen layer, a vector of values is formed, which corresponds to a specified cluster, which is then transmitted to the input of the perceptron network which, based on approximation of the complex nonlinear relationship between values of the gradient of the local temperature field at the prediction point and the level of water, calculates predicted values of the gradient of the local temperature field and the level of water.EFFECT: high address accuracy, reduced errors, reduced labour input in the prediction process.5 cl, 4 dwg, 2 tbl, 1 ex |
