پيش بيني بارش ايستگاه بابلسر توسط مدل بهينه يافته موجك ماشين آموزش نيرومند

Simulation of Long-Term Rainfall in Babolsar City by Using Optimized Wavelet-Extreme Learning Machine

در اين مطالعه، بارندگي دراز مدت شهر بابلسر توسط يك مدل هوش مصنوعي بهينه يافته شبيه‌سازي شد. براي اينكار، ماشين آموزش نيرومند (ELM) و تبديل موجك (wavelet transform) با همديگر تركيب شدند. لازم به ذكر است كه مقادير بارندگي‌ها به‌صورت ماهيانه از سال 1951 تا 2019 بكار گرفته شدند كه 70 درصد آنها براي آموزش اين مدل هوش مصنوعي و 30 درصد باقيمانده براي آزمون آن استفاده گرديد. در ابتدا، توابع فعال‌سازي مدل ماشين آموزش نيرومند مورد بررسي قرار گرفتند كه بهترين آن شامل تابع فعال‌سازي sigmoid انتخاب شد. همچنين، تاخيرهاي داده‌هاي سري زماني با استفاده از تابع خود همبستگي معرفي شدند كه با استفاده از اين تاخيرها، چهار مدل ماشين آموزش نيرومند تعريف گرديد. با اجراي يك تحليل حساسيت، مدل برتر ELM معرفي شد. مقادير ضريب همبستگي (R)، شاخص عملكرد (VAF) و شاخص پراكندگي (SI) براي مدل برتر ELM به‌ترتيب مساوي با 0/524، 27/064 و 0/819 محاسبه شدند. علاوه بر اين، موجك‌هاي مادر مختلف مورد بررسي قرار گرفتند كه dmey به‌عنوان بهترين موجك مادر انتخاب شد. تبديل موجك دقت مدل‌سازي را به شكل قابل توجهي افزايش داد. به‌عنوان مثال، شاخص عملكرد مدل تركيبي WELM مساوي با 461/86 بدست آمد. لازم به ذكر است كه مدل تركيبي نيز براي سه سطح تجزيه مختلف مورد بررسي قرار گرفت كه بهترين سطح تجزيه مدل تركيبي معرفي گرديد. لازم به ذكر است كه تاخيرهاي شماره (t-1) و (t-12) به‌عنوان موثرترين تاخيرهاي داده‌هاي سري زماني شناسايي شدند.

Simulation of rainfall is considered as one of the most important issues of hydrologic field. Additionally, the artificial intelligence (AI) models own numerous privileges including acceptable accuracy, appropriate flexibility, and high speed. In this study, the long-term rainfall in Babolsar Station was simulated by means of an optimized AI model. To do this, the extreme learning machine (ELM) and the wavelet transform (WT) were combined. It should be stated that the monthly rainfall values from 1951 to 2019 were applied, meaning that 70% of the observed values were employed to training the AI models and 30% of rest were utilized to testing these models. Firstly, the activation functions of the ELM models were evaluated; as a result, the sigmoid was chosen as the best activation function. Moreover, the lags of time series were introduced using the autocorrelation function (ACF) that four ELM models were defined through those identified lags. By performing a sensitivity analysis, the superior ELM model was introduced. The values of correlation coefficient (R), variance accounted for (VAF), and scatter index (SI) for the ELM model were respectively computed to be 0.524, 27.064, and 0.819. Furthermore, different mother wavelets were examined and the “dmey” was opted as the best mother wavelet. The wavelet transform enhanced the accuracy of the simulations significantly. For instance, the VAF index for the hybrid WELM model equaled to 86.461. It is noteworthy that the hybrid model was evaluated for different decomposition levels (DL) and then the best one was detected. Also, the (t-1) and (t-12) lags were identified as the most effective input lags.