Multiresolution neural networks for electricity market forecasting: addressing complexity with q -wavelet methods
提出一种集成q-最大重叠离散小波变换的前馈神经网络框架，用于预测澳大利亚电力现货价格，通过分离规则与不规则分量提升计算效率，预测精度比传统模型提高2到20倍。
Abstract This article introduces a multiresolution neural network framework for forecasting electricity spot prices, exemplified by the Australian market’s mature and intricate structure. Central to our methodology is the q-Maximal Overlap Discrete Wavelet Transform, an advanced preprocessing technique integrated into a feedforward neural network (FFNN) to capture long-memory components and enhance feature representation. Unlike traditional wavelet-based methods, our approach consolidates processed sub-signals into regular and irregular components, modelled separately using specialized FFNNs to improve computational efficiency. The inclusion of dummy variables, such as Daylight Saving Time, further supports the integration of exogenous factors, enhancing robustness and interpretability. The proposal achieves improvements of up to 20-fold over a conventional single-hidden-layer FFNN, two- to 15-fold over univariate state-of-the-art models, and two- to more than ninefold over comparable bivariate models, underscoring its accuracy and stability. These results establish the framework as a decision-support tool for navigating the uncertainty and heterogeneity of modern energy systems while advancing operational research methodologies and promoting efficient, data-driven strategies for energy transitions.