ارزیابی کارایی مدل‌ LSTM در پیش‌بینی جریان روزانه ورودی به مخازن سدها

نوع مقاله : مقاله پژوهشی

نویسندگان

• معصومه زینعلی ¹
• امید بزرگ حداد ¹
• مهدی یاسی ²
• حسین علیزاده ³

¹ گروه مهندسی آبیاری و آبادانی، دانشکدگان کشاورزی و منابع طبیعی دانشگاه تهران، کرج، ایران.

² استاد، گروه مهندسی آبیاری و آبادانی، دانشکدگان کشاورزی و منابع طبیعی، دانشگاه تهران، کرج، ایران

³ گروه آب و محیط زیست، دانشکده مهندسی عمران، دانشگاه علم و صنعت ایران، تهران، ایران

چکیده

کلیدواژه‌ها

• افق پیش‌بینی بلندمدت
• پیش‌بینی جریان روزانه مخزن
• مدیریت سیلاب
• یادگیری ماشین

موضوعات

• منابع آب

عنوان مقاله [English]

Performance evaluation of the LSTM Model forecasting daily inflow into dams reservoirs

نویسندگان [English]

• masoumeh zeinalie ¹
• omid bozorg haddad ¹
• Mehdi Yasi ²
• Hosein Alizadeh ³

کلیدواژه‌ها [English]

• daily reservoir inflow forecast
• flood management
• long lead-time forecast
• machine learning

EXTENDED ABSTRACT

Background and purpose

Real-time forecasting of daily inflows to reservoirs with a prediction horizon that extends several steps into the future is crucial for water resource planning and management. Despite numerous studies on inflow prediction using machine learning methods, few have investigated the predictive capabilities of these approaches with long lead time (several steps ahead) or gained insights through systematic comparisons of model predictive performance in the short term. Therefore, in the present study, an advanced type of recurrent neural network called the LSTM model has been introduced to predict the daily inflow to dam reservoirs, addressing the flaws and defects of previously used methods. LSTM is a type of deep learning neural network that uses complex and combined functions instead of simple functions to regulate and strengthen short-term memory. By optimally setting modeling parameters such as the number of neurons in each layer, the number of iterations, and the number of categories, the LSTM model has a high potential for long-term prediction. Therefore, the purpose of this research is to evaluate the efficiency of and compare the machine learning models of the Adaptive Neuro-Fuzzy Inference System (ANFIS) and the Long Short-Term Memory (LSTM) network in predicting the daily inflow to dam reservoirs. Subsequently, the inflow to the Seimareh reservoir is modeled using these methods, with a particular focus on long-term forecasting (7 days ahead). Finally, the superior model is used to predict the inflow to the reservoir under flood conditions.

Materials and methods

The stages of the research include three main parts: data collection, data pre-processing, and modeling. First, the effectiveness of the Adaptive Neuro-Fuzzy Inference System (ANFIS) model and the Long Short-Term Memory (LSTM) network was evaluated to predict the daily inflow to the Seimareh reservoir in the Karkheh basin, for a forecast horizon of the next 7 days. For this purpose, daily data on precipitation, temperature, and inflow to Seimareh from 2012 to 2018 were used for modeling. Then, using the superior model that is selected based on the evaluation indicators, the daily inflow to the Seimareh reservoir was forecasted for the next 7 days during the flood of 2019.

Findings

In this section, the evaluation indices of the Nash coefficient, correlation coefficient, root mean square error, and bias value for selected models in the forecast horizons of one to seven days are presented. For the ANFIS model, the further the prediction horizon extends, the lower the value of the explanation coefficient, leading to decreased performance in simulating abnormal flow values and an increase in the number of simulated outliers. Regarding the flow simulated with the LSTM model in this case study, it cannot be said that the accuracy of the simulation decreases as the horizon extends. For instance, in the forecast horizon of the next seven days, the flow simulation is better than that for the next six days, and the performance of the model for one, two, and five days is better than for the other horizons. In general, the LSTM model has performed well in simulating abnormal flow values, and the number of simulated outliers is not significant across all forecast horizons.

Conclusion

The results showed that the performance of the LSTM model is better than that of ANFIS in daily forecasting several steps ahead. The maximum and minimum values of the Nash coefficient in the forecast horizon for the next seven days were 0.971 and 0.628, respectively, for the LSTM model, and 0.858 and 0.393 for the ANFIS model. The optimal setting of parameters, including the number of neurons in each layer, the number of iterations, and the number of categories in the LSTM model, is key to the model's high potential to predict flow up to the forecast horizon of the next seven days. Finally, the performance of the LSTM model in predicting the inflow to Seimareh during the flood of 2019 was evaluated, and it was found to predict the flood flows with acceptable accuracy up to the forecast horizon of the next 7 days. This study can provide valuable insights to water resource managers for planning and managing daily reservoir discharge in real-time operations. findings show that the LSTM model is suitable for forecasting daily flow. Therefore, it can be used to aid strategic decisions in the management of water resources, especially under flood conditions.

Author Contributions

All authors contributed equally to the conceptualization of the article and writing of the original and subsequent drafts.

Data Availability Statement

If the study did not report any data, you might add “Not applicable” here.

Acknowledgements

The authors would like to thank all participants of the present study.

Ethical considerations

The authors avoided data fabrication, falsification, plagiarism, and misconduct.

Conflict of interest

The author declares no conflict of interest.