پديد آورندگان :
نشاط، نجمه دانشگاه ميبد - گروه مهندسي صنايع (سيستم‌ها) , سرداري زارچي، محسن دانشگاه صنعتي شريف - گروه مهندسي كامپيوتر , محلوجي، هاشم دانشگاه صنعتي شريف - دانشكده مهندسي صنايع
كليدواژه :
شبكه‌هاي عصبي عميق , پيش‌بيني , شبكه‌هاي تمام‌متصل , مدل عميق تركيبي , بار اوج مصرفي
چكيده فارسي :
اين مطالعه به بررسي كارايي پيكره‌بندي مختلف شبكه‌هاي يادگيري عميق (رويكرد برتر در مدل‌سازي و تخمين سري‌هاي زماني اقتصادي ـ اجتماعي) در حوزه‌ي پيش‌بيني مي‌پردازد. در اين مطالعه به‌منظور ملموس‌سازي رويكرد پيشنهادي از مدل‌سازي و پيش‌بيني اوج بار مصرفي خانگي در قالب موردكاوي استفاده شده است. نتايج حاكي از برتري توپولوژي شبكه تركيبي از تمام متصل و بازگشتي بود كه اين برتري با توجه به ماهيت غيرخطي و پيچيده، وابستگي‌هاي قوي به داده‌هاي دوره‌هاي قبلي و همچنين وجود درجات متفاوتي از تأخير در متغيرهاي برون‌زاي مسئله كاملاً توجيه‌پذير است. نظر به اين‌كه در اين مدل متغيرهاي برون زاي مدل) نماينده‌ي شرايط مختلف جوي (و متغيرهاي مصنوعي) نماينده‌ي شرايط مختلف زماني( نيز لحاظ شده است، از استواري قابل قبولي نسبت به مدل‌هاي ارائه شده در مطالعات قبلي برخوردار است.
چكيده لاتين :
This study examines the efficiency of various topologies of deep learning networks (a superior approach to modeling and fitting socio-economic time series) in load demand forecasting using the data collected from a four-year period of households in Kurdistan City, Iran. Since the consumption pattern is a nonlinear and complex curve with a strong delayed dependency pattern, its prediction is not accurate by conventional statistical methods and the error reduction of this prediction has a significant effect on reducing production costs, unwanted squandering and fines. In this study, full-connected, recurrent, and also hybrid of them were investigated using the mean efficiency of absolute error percentage and mean square error index. When the input of the neural network was in the form of tensor, designing the structure of the deep neural network would be straightforward. In this case, the network can be implemented with a linear stack of layers sequentially. Although the sequential the sequential model is so common, it is inflexible when the input data is not in the form of tensor, e.g., Figure 4. Besides, in a forecasting model, each determinant might need a different type of neural networks such as CNN, LSTM or GRU. To overcome this challenge, we innovatively proposed parallel deep branches in our framework to represent the history of each determinant individually. The parallel branches process their determinants by using RNN and Dense networks. Then, the branches were merged together through concatenated and dense layers. The results indicated the superiority of the network topology as a combination of all connected and reciprocating models for modeling and predicting
consumption. This superiority, due to the nonlinear nature of complexity, the strong attachment to the data of previous periods, and the existence of different degrees of delay in the exogenous variables of the problem can be fully justified. Considering that for excited peak load prediction, exogenous variables of the model (representing different atmospheric conditions) and artificial variables are included, this model has acceptable stability, compared to the models presented in previous studies.
