Deep Fusion Predictive Modeling Based on Noise Analysis

Noise-containing measurement data is widespread in actual systems, and its accurate prediction is of great significance to human production and life. However, the existence of noise and the non-linear characteristics of the system interfere with the modeling of the prediction model, so the accuracy of the data prediction of the actual time series has never reached human expectations. This paper proposes a hybrid prediction model based on wavelet threshold analysis, Bayesian optimization method and deep learning. First, the prediction model smooths the noise of the original data through wavelet threshold analysis. Secondly, the smoothed data is decomposed by wavelet decomposition to reduce the complexity of the data. Then use Gated Recurrent Unit (GRU) as a sub-predictor for training for each component. The overall network uses Bayesian Sequence Model Optimization (SMBO) to select hyperparameters. Finally, the prediction results of each sub-predictor are added to obtain the final prediction result. The research verifies the effectiveness of this method by comparing the prediction performance of other networks in Beijing PM2.5 data.