Deep learning neural networks have many applications in various topics such as prediction, classification and pattern recognition. LSTM neural network is a deep learning method that can be used to detect weather conditions. In the first phase, the crow's learning algorithm is used to select the feature in the first layer of the LSTM network. The evaluations showed that the accuracy of the proposed method is 96.92% and without using the feature selection, it is about 93.21%. The evaluations show that the accuracy of the proposed method for forecasting the weather condition is higher than LSTM and MLP methods.

Keywords: deep learning, LSTM network, crow learning algorithm, weather forecasting



1. Introduction

Meta-heuristic algorithms are problem solving methods that are modeled on the events in nature or the behavior of living beings so that optimal solutions can be extracted. Collective intelligence algorithms [1] are a kind of meta-heuristic algorithms that are modeled on the behavior of living beings that live in a group and social life, such as hunting behavior, hyena optimization algorithm, whale optimization algorithm, etc. Is. Meta-heuristic algorithms can be divided into different categories based on the method of problem solving, one of which is shown in the research [2] in 2020 according to the diagram in Figure (1) and can be seen. Meta-heuristic algorithms are divided into 4 different groups and categories based on their performance:



Figure 1: Classification of meta-heuristic algorithms into different categories [2]

Meta-heuristic algorithms are used in various fields, one of which is the optimization of machine learning and deep learning parameters. One of the applications of machine learning and deep learning is in weather forecasting. In this article, to improve the accuracy of the LSTM network, the optimization of important features using the learning method in crows has been used.

2. LSTM learning network

In the short term, long memory neural networks are actually a type of recurrent neural networks [3].

In the LSTM network, with the help of the sigmoid function that is applied element by element, the input, forgetting and output gate layers produce vectors whose all dimensions are between zero and one or close to both. The general structure of LSTM deep learning neural network is as shown in Figure (2):

Figure 2. The structure of long memory networks, in the short term

3. proposed model



Figure (3). the framework of the proposed method is shown.



The evaluation or minimization function of the following two factors shows how well a feature vector has competence:

• Average prediction error with neural network

• Number of features selected



The calculation of error index E is as follows:

the population of crows is stored in a matrix:

Each crow needs to remember the most optimal position:

In the crow learning algorithm, there are two phases of horizontal and vertical learning. Vertical learning from parents is horizontal learning from brothers and sisters.

It is used to select a sister or vector randomly.

6 k= 3+[rand×(i-3)] & i≥3

In the crow's learning algorithm, the probability of receiving a reward for crows is equal to Rpprob.



lf is the value of the learning factor in crows.



Amount of reward for crows:



Reinforcement of learning for parents is used in the Crow algorithm as follows.



It is used to search for food with the stealing mechanism as follows:



4-Implementation and analysis

4-1-Implementation parameters



Table (1): Implementation parameters of the proposed method



Figure 4: LSTM implementation parameters in the proposed method



4-2-Evaluation indices

One of the important indicators for predicting weather conditions is the mean squared error MSE index, and to evaluate the proposed method, you can use the classification and prediction indicators of accuracy, recall and accuracy:



4-3- Analysis of the proposed method

In the diagram of figure (5), the prediction error in the feature selection phase in combination with the neural network is shown, and in figure (6), the output of LSTM deep learning in weather forecasting is depicted.



Figure 5: Reduction of prediction error in feature selection phase with 10 iterations

Figure 6: Reducing the prediction error in the classification phase with LSTM



Table 2: Average prediction indices of the proposed method



Figure 7: Comparison of the MSE error of the proposed method with prediction methods

Figure 8: Comparison of the accuracy of the proposed method with prediction methods

Figure 9: Comparison of recall of the proposed method with prediction methods

Figure 10: Comparing the precision of the proposed method with prediction methods

Figure 11: Comparison of the accuracy index of the proposed method in weather forecasting



5. Conclusion

LSTM network is a deep learning method that can be used to predict weather conditions. In the proposed method to increase the prediction accuracy of LSTM neural network, intelligent feature selection is used using a combination of crow learning algorithm and crow search. Experiments showed that the proposed method has an accuracy of 96.92%, a sensitivity of 95.82%, and an accuracy of 96.34%, and it is more accurate for predicting weather conditions than multilayer neural network, recurrent neural network, and LSTM method.



Reference:

[1] Xue, J., & Shen, B. (2020). A novel swarm intelligence optimization approach: sparrow search algorithm. Systems Science & Control Engineering, 8(1), 22-34.

[2] Li, S., Chen, H., Wang, M., Heidari, A. A., & Mirjalili, S. (2020). Slime mould algorithm: A new method for stochastic optimization. Future Generation Computer Systems.

[3] Fu, J., Chu, J., Guo, P., & Chen, Z. (2019). Condition monitoring of wind turbine gearbox bearing based on deep learning model. Ieee Access, 7, 57078-57087.

[4] Sampaio, P. S., Almeida, A. S., & Brites, C. M. (2021). Use of artificial neural network model for rice quality prediction based on grain physical parameters. Foods, 10(12), 3016.