Forecasting technique of radio information system characteristics for optimal technical condition of the space control system components

Abstract

The paper proposes a forecasting technique of radio information system characteristics for optimal technical condition of space control system components, using a recurrent neural network and information tools from an early-warning radar stations group and Russian Aerospace Forces optical-electronic stations when determining the optimal technical condition of the Space Monitoring System (SMS) components. Analyzing values of time series by a recurrent neural network LSTM-based model with various parameters, we propose an algorithm for solving problem of forecasting characteristics of a radio information complex. The network architecture is a two-layer network with one layer of LSTM cells and one fully connected layer. There are results of experimental forecasting by pre-training the recurrent neural network (RNN) model sampling 1080 normalized measurements of a conditional sensor of 30 minutes apart. The training sample is made of conditional sensor normalized values for 9 hours continuous operation. The algorithm for the model training and creation is implemented in Python language using open source library TensorFlow. Experiments were performed with different values of the forecast horizon and the length of the input sequence of time series values. We obtained: the total average error for each combination of parameters, the dependence of the average forecast error on the used parameters and the dependence of the average (for all forecast values) forecast error on the period of the analyzed sensor values.