عنوان مقاله [English]
In today's world, rapid and sustainable development is the ultimate goal of all countries. The main constraint to sustainable development is the climatic conditions of countries, including rainfall, which is most correlated with natural disasters. One of the methods used for prediction is deep neural networks. The main purpose of this study is to increase the accuracy of hourly rainfall forecasting in Khorasan Razavi region using one of the deep neural network methods. In this study, we present a deep neural network architecture using the Ron Neuron-Based Random Neural Decomposition Stochastic (RSDSAE) method for short-term precipitation prediction. In order to improve the accuracy, rough neural networks (RNNs) contribute to the rainfall uncertainty and to improve the speed and accuracy, the combination of the sparse algorithm with the above model has been used and We also used the predicted output data output of the WRF model, and tests are calculated on rain data and by two criteria of RMSE, MAE and with five self-esteem, respectively, 0.7912 and 0.7662, and has been able to better performance than the model (RSDAE) of itself Show.
Keywords: Deep Neural Network, Precipitation Prediction, stacked denoising sparse autoencoder
Rainfall is a climatic factor that affects many human activities such as agriculture, construction, electricity generation, etc. Therefore, having a proper method for predicting rainfall can take preventive and mitigating measures for natural disasters. Over the past few years, deep neural networks (DNNs) have been used as a successful mechanism for solving complex problems in areas such as machine vision, image recognition, etc. (Hern´andez, et al., 2016). Deep neural network is a set of multi-layered architectures that have been trained using unsupervised algorithms and have challenges in training deep neural networks such as slow network training, overprocessing, and so on. Most existing neural network models have three drawbacks: (Khodayar, et al., 2017)
1- Most of the architecture is shallow.
2. Some methods require handmade engineering features that are tedious.
3- Most methods do not have direct knowledge about rainfall uncertainty.
In this study, to solve problems one and two, we combined the proposed method of stacked denoising sparse autoencoder (SDSAE) and to solve problem three, we combined the proposed method with Neuron Rough.
2. Materials and methods:
2-1. Data and area of study:
In this study, the study area is Khorasan Razavi province and the output of networked data is the WRF model, which is a mid-scale regional forecasting model. We considered eight days of regional precipitation hourly data as target variable data for training and testing algorithms. We prepared the data used from the output of two regional forecast model implementations (mid-scale atmospheric forecast model) on 26 and 28 October 2018 for each implementation with 4 days forecast.
2-2. Steps to perform the proposed method:
The proposed method has five main steps:
1) Receive WRF output prediction data,
2) Apply noise to input data.
3) Validation of deep network parameters.
4) Unsupervised learning with RSDAE to build deep multilayer network.
5) Learning by monitoring (fine tuning) by BP algorithm with SGD method
2-3. Evaluation criteria
To evaluate the proposed method SRDSAE and RSDAE, we use two criteria, mean squared error (RMSE), absolute mean error (MAE).
3. Research results:
Predicting rainfall is one of the most important and fundamental challenges for researchers. Properly predicting rainfall improves the lives of the general public and even the proper planning of governments for the correct use of rainfall. Various algorithms and methods have been proposed for forecasting issues in recent years. Deep learning methods are one of the most popular methods for predictive problems. In this research, we have used the stacked denoising autoencoder method to predict rainfall and combine this method with neuron-rough and sparse algorithm to increase the accuracy of the forecast. In this research, we applied the proposed method on the output data of the WRF mid-scale forecast model in Khorasan Razavi province.
The proposed RSDSAE(Rough Neuron Based stacked denoising sparse autoencoder) method and RSDAE (Rough Neuron Based stacked denoising autoencoder) method have been compared and analyzed for better comparison in precipitation forecasting based on two evaluation parameters RMSE and MAE. The proposed method has been able to reduce the forecast error by reducing the RSDAE network and subsequently reduce the RMSE and MAE criteria, or in other words, improve the rainfall forecast.
1- Hern´Andez, E. et al., 2016. Rainfall Prediction: A Deep Learning Approach. in International Publishing Switzerland , Springer, pp. HAIS, LNAI 9648, pp. 151–162.
2- Khodayar, M., Kaynak, O. & Khodayar, M. E., 2017. Rough Deep Neural Architecture for Short-term Wind Speed Forecasting. IEEE, pp. 1551-3203 (c).
3- Kumar, V., Nandi, G. & Kala, R., 2014. Static Hand Gesture Recognition using Stacked Denoising Sparse Autoencoders. Robotics and Artificial Intelligence Laboratory.IEEE, pp. 978-1-4799-5173-4).