Establishing the energy consumption prediction model of aluminum electrolysis process by genetically optimizing wavelet neural network

Nowadays, it is very popular to employ genetic algorithm (GA) and its improved strategies to optimize neural networks (i.e., WNN) to solve the modeling problems of aluminum electrolysis manufacturing system (AEMS). However, the traditional GA only focuses on restraining the infinite growth of the optimal species without reducing the similarity among the remaining excellent individuals when using the exclusion operator. Additionally, when performing arithmetic crossover or Cauchy mutation, a functional operator that conforms to the law of evolution is not constructed to generate proportional coefficients, which seriously restricted the exploitation of the hidden potential in genetic algorithms. To solve the above problems, this paper adopts three new methods to explore the performance enhancement of genetic algorithms (EGA). First, the mean Hamming distance (H-Mean) metric is designed to measure the spatial dispersion of individuals to alleviate selection pressure. Second, arithmetic crossover with transformation of the sigmoid-based function is developed to dynamically adjust the exchange proportion of offspring. Third, an adaptive scale coefficient is introduced into the Gauss-Cauchy mutation, which can regulate the mutation step size in real time and search accuracy for individuals in the population. Finally, the EGA solver is employed to deeply mine the optimal initial parameters of wavelet neural network (EGAWNN). Moreover, the paper provides the algorithm performance test, convergence analysis and significance test. The experimental results reveal that the EGAWNN model outperforms other relevant wavelet-based forecasting models, where the RMSE in test sets based on EGAWNN is 305.72 smaller than other seven algorithms.