Time Series Forecasting Based on Reptile Search RSA Optimized LSTM

0 Introduction

The time series prediction method based on LSTM is simple and effective.The emergence of LSTM provides a new research direction for time series forecasting.However, like most network models, LSTM performance is affected by its hyperparameter settings.To this end, this paper uses the Reptile Search Algorithm (RSA) to optimize the LSTM network hyperparameters, and establishes the RSA-LSTM model. The example verification shows that the prediction effect of the RSA-LSTM model is significantly improved.

1 Principles

1.1 LSTM principle

1.2 Reptile Search Algorithm

Reptile Search Algorithm (RSA) is a new nature-inspired meta-heuristic optimizer proposed by Laith Abailigah et al. in 2020. Its inspiration comes from the social behavior of crocodiles in nature. It mainly includesTwo main mechanics: encirclement, and hunting.The mathematical model of these two mechanisms is established, that is, the RSA algorithm is proposed.The RSA algorithm is a population-based gradient-free method that can be used to solve complex or simple optimization problems with specific constraints.For specific theory, please refer to Reptile Search Algorithm (RSA): A novel nature-inspired meta-heuristic optimizer

1.3 SMA Optimization LSTM Principle

With minimizing the error of the LSTM network as the fitness function, the role of RSA is to try to find a set of optimal hyperparameters to minimize the network error.The main hyperparameters of LSTM in this paper are: learning rate lr, batchsize, training times K, and the number of nodes in the two hidden layers L1 and L2.

2 Code Implementation

Based on MATLAB2020b, model building and optimization.The data structure is a time series. We use the value of the first n times as input and the value of time n+1 as the output to carry out rolling modeling.

2.1 LSTM results

2.2 RSA-LSTM results

The following figure shows the change curve of different hyperparameters:

The results of the RSA-LSTM model established using the above optimal parameters are:

2.3 Method comparison

3 Conclusion

It can be seen from the above analysis that the optimized LSTM has better accuracy.