Optimized deep networks structure to improve the accuracy of estimator algorithm in deep networks learning

The present study considers decreasing prediction error for various types of time series and the uncertainty in estimation parameters, improving the structure of the Deep Neural Network (DNN), and increasing response speed in the proposed neural network method. Additionally, the competitive performance and the collaboration among the neurons of the DNN are also increased. The selected data is related to weather prediction for Qeshm, which has suitable weather conditions for our study, spanning from 2016 onwards. Tn this study, to analyze the prediction issue of power consumption of domestic expenses in the indefinite and severe fluctuation mode, we decided to combine the two methods of Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN). For the training of the deep network, the Rack Propagation (RP) algorithm is used. The results indicate that Gated Recurrent Unit (GRU) networks compared to other models (MultiLayer Perceptrons (MLP), CNN, and DNN) produce more realistic results, and also twoway networks obtained better results on test data compared to LSTM networks. Root Mean Square Error (RMSE) prediction is more realistic than the LSTM model on test and training data against significant data. A GRU network has two gates of rt readjust and Zt forgetting, which helps to ensure that long-term dependencies of gradient fading will not occur. (c) 2024 Sharif University of Technology. All rights reserved.