PPTformer: A novel hybrid model for enhanced long-term time series forecasting with extreme value focus

Long-term time series forecasting (LTSF) is crucial for strategic planning and the development of early warning systems. While recent models have shown promise, they struggle to capture long-term dependencies and forecast critical points like peaks and valleys. To address this, we present PPTformer, a novel model that combines ensemble learning and specialized modules to enhance the precision of extreme value forecasting. PPTformer includes a peak-valley forecasting module for recognizing and predicting extreme features, a partition and detrending preprocessing module to isolate complex patterns and highlights intrinsic properties to the forefront, and a time-lag attention convolutional encoder module adept at capturing long-term dependencies. Furthermore, PPTformer breaks away from the traditional encoder-decoder paradigm by adopting a streamlined architecture that replaces the decoder with a linear projection layer, effectively overcoming the limitations of traditional models. Extensive tests on 9 real-world datasets show that PPTformer outperforms both traditional and state-of-the-art models on 6 datasets, and performs second best on the remaining datasets, demonstrating its effectiveness for practical real-world LTSF applications.