Time sequence prediction-oriented drift pulse neural network construction method and application thereof

Abstract

The invention belongs to the technical field of time sequence prediction, and particularly relates to a time sequence prediction-oriented drift pulse neural network construction method and application thereof, and the method comprises a new coding method which comprises the steps: converting an original time sequence into time when neurons give out pulses for the first time; besides, the network input layer is responsible for issuing pulses at corresponding moments according to the coded z-domain time sequence data Z (xt), the cell state Ct-1 at the previous moment and the hidden state Z (ht-1); the hidden layer comprises a forgetting gate, an input gate and an output gate, and the input of each gate is a pulse corresponding to Z (xt) and Z (ht-1); the pulse generated by the forgetting gate is used for realizing a function of selectively forgetting Ct-1; the pulse generated by the input gate is used for updating the cell state at the current moment; the forgetting gate and the input gate jointly realize the long-time memory ability of the time sequence; the pulse generated by the output gate corresponds to the hidden state at the current moment. According to the invention, the problem of time delay between input and output neurons can be effectively solved, and the long-time memory capability is realized.

Description

technical field [0001] The invention belongs to the technical field of time series prediction, and more particularly, relates to a method for constructing a drift pulse neural network for time series prediction and its application. Background technique [0002] Time series forecasting is of great significance in scientific development, technological breakthroughs, and practical applications. Taking the power system as an example, time series forecasting includes power generation forecasting, electricity price forecasting, and load forecasting. This type of time series prediction is beneficial to environmental protection and economic benefits as a whole, to power dispatch and safe operation on the grid side, and to energy use and behavioral intervention on the user side, so it has important research value. [0003] Long-short-term memory neural network is the mainstream model for time series forecasting. Compared with artificial neural networks, recurrent neural networks, a...

AI Technical Summary

Problems solved by technology

The first is frequency coding, which realizes coding by calculating pulse counts and ratios. The defect is that discrete pulse counts and pulse ratios will make network training difficult; the second is time coding, which uses the time pair of pulse neurons to emit pulses. Data is encoded, and its defect is that when encoding time series, there is an obvious delay in the time interval between the input neuron and the output neuron.

In addition, due to the lack of memory ability of spiking neural network, especially long-term memory ability, its prediction effect on time series is only better than that of traditional artificial neural network of the same size

Images