A pattern classification methodology for interval forecasts of short-term electricity prices based on hybrid deep neural networks: A comparative analysis

Abstract

Precisely identifying multidimensional trends and hidden characteristics that relate to short-term electricity price fluctuations and providing reliable predictions of future trends are difficult tasks. Instead of utilizing conventional regression-based forecasting methods, we present a novel methodology to address the problem of obtaining reliable forecasts from a pattern classification standpoint. Given that an attention mechanism is better able to capture global characteristics, a multi-head self-attention (MHSA) mechanism is adopted to extract features at long time scales more efficiently and ensure that long-term dependencies can be captured. On this basis, a new hybrid framework composed of nested long short-term memory (NLSTM), the MHSA mechanism combined with a convolutional neural network (MHSAC), and a feature space identification approach is established for robust interval forecasts. To verify the performance of our framework and demonstrate its application potential, the proposed classifier is compared to benchmarks under various scenarios (8 different input dimensions and 25 different input sizes) in terms of different performance criteria. The results indicate that our framework can be used as a valid alternative for electricity price forecasting, and it achieves satisfactory forecasting accuracy with different input dimensions. Furthermore, compared with regression-based models and price spike forecasting cases, our framework is more effective than the benchmarks. The findings also suggest that appropriate feature selection is more conducive to improving model forecasting accuracy than blindly increasing the dimensions of the input data, and the proposed framework that incorporates the MHSA mechanism is also propitious for further improving the efficiency of electricity price forecasting.