Recurrent transform learning

Deep Learnign Techniques for Demand Forecasting

The aim of this study is to categorize research on the applications of deep learning techniques in demand forecasting and suggest further research directions. This study is based upon 56 papers published between 2017 and April 2021 in international peer-reviewed elite journals. Primary objective of this paper is to identify the major problem domains in demand forecasting; hence authors conduct a review of literatures which utilizes deep learning techniques for demand forecasting and proposed directions for future research. After identifying the objective a subject scrutiny of the important papers are done based on the publication quality. These identifications make additions to demand forecasting research in the resulting manner. For accomplishing this task, first, the authors classified the literature into nine major problem domains based on different issues discussed in the literature. Second, the literatures are classified based on different deep leaning techniques used for solving the problem of demand forecasting. Third, seven research propositions provided for future research.

An adaptive backpropagation algorithm for long-term electricity load forecasting

Artificial Neural Networks (ANNs) have been widely used to determine future demand for power in the short, medium, and long terms. However, research has identified that ANNs could cause inaccurate predictions of load when used for long-term forecasting. This inaccuracy is attributed to insufficient training data and increased accumulated errors, especially in long-term estimations. This study develops an improved ANN model with an Adaptive Backpropagation Algorithm (ABPA) for best practice in the forecasting long-term load demand of electricity. The ABPA includes proposing new forecasting formulations that adjust/adapt forecast values, so it takes into consideration the deviation between trained and future input datasets' different behaviours. The architecture of the Multi-Layer Perceptron (MLP) model, along with its traditional Backpropagation Algorithm (BPA), is used as a baseline for the proposed development. The forecasting formula is further improved by introducing adjustment factors to smooth out behavioural differences between the trained and new/future datasets. A computational study based on actual monthly electricity consumption inputs from 2011 to 2020, provided by the Iraqi Ministry of Electricity, is conducted to verify the proposed adaptive algorithm's performance. Different types of energy consumption and the electricity cut period (unsatisfied demand) factor are also considered in this study as vital factors. The developed ANN model, including its proposed ABPA, is then compared with traditional and popular prediction techniques such as regression and other advanced machine learning approaches, including Recurrent Neural Networks (RNNs), to justify its superiority amongst them. The results reveal that the most accurate long-term forecasts with the minimum Mean Squared Error (MSE) and Mean Absolute Percentage Error (MAPE) values of (1.195.650) and (0.045), respectively, are successfully achieved by applying the proposed ABPA. It can be concluded that the proposed ABPA, including the adjustment factor, enables traditional ANN techniques to be efficiently used for long-term forecasting of electricity load demand.