基于元模式挖掘与提示引导的时空预测方法

Spatio-temporal prediction plays a critically indispensable role in numerous real-world scenarios, such as traffic forecasting, urban planning, environmental monitoring, and disaster prevention; however, a significant limitation persists in existing spatio-temporal prediction methodologies. These approaches tend to focus excessively on global average characteristics while systematically neglecting the inherent heterogeneity embedded within spatio-temporal data structures. This oversight makes it fundamentally challenging to effectively capture nuanced local differences and dynamic temporal variations in spatio-temporal patterns across diverse spatial regions or evolving time periods, ultimately culminating in suboptimal prediction performance that fails to address real-world complexity. To address this pervasive issue, this paper proposes a novel spatio-temporal prediction method named STMP2G (Spatio-Tcmporal Meta-Pattcrn Prompt Guidance), which innovativcly integrates meta-pattern mining with prompt guidance mechanisms. The framework employs a deliberately simplified multi-layer pcrccptron (MLP) as its core prediction network, thereby prioritizing adaptability over architectural complexity. Crucially, STMP2G captures spatio-temporal heterogeneity by dynamically generating specialized spatio-temporal prompts. These prompts undergo enhancement through a dedicated self-supervised learning phase designed to mine universally applicable meta-patterns from rich historical data repositories. Ultimately, the refined prompts guide the entire spatio-temporal modeling process, strategically overcoming heterogeneity interference to achieve both accurate and robust predictions even in highly non-stationary environments. Specifically, the method first constructs a spatio-temporal prompt generation module. This component generates multi-scale temporal prompts spanning hourly, daily, and weekly granularities using temporal prior embedding combined with a memory network to comprehensively capture temporal heterogeneity. Simultaneously, it produces semantic spatial prompts via learnable spatial embedding matrices optimized through representation learning objectives, thereby encoding region-specific attributes reflecting spatial heterogeneity. To model intricate interdependencies, the module constructs a high-order spatio-temporal interaction tensor and achieves deep fusion of spatial and temporal prompts using advanced inverse tensor decomposition techniques, ensuring cohesive cross-dimensional feature integration. Subsequently, STMP2G incorporates a self-supervised meta-pattern mining module. This stage employs contrastive self-supervised learning to discover transferable, discriminative meta-patterns from extensive historical sequences within a unified feature space. These foundational patterns once identified have their parameters permanently frozen and are systematically archived within a structured meta-pattern library, creating a reusable knowledge base for efficient real-time retrieval during inference. Finally, the architecture features a spatio-temporal prompt guidance module. During prediction, this component retrieves the top-k most relevant meta-patterns from the meta-pattern library based on similarity to the input spatio-temporal sequence, u-sing these to contextually supplement and enhance the initial prompts. The augmented prompts are then mapped to the feature space of each MLP layer via lightweight, layer-specific adapters. Through Hadamard product operations, the module achieves fine-grained, adaptive enhancement of evolving spatio-temporal representations, thereby guiding the modeling process to dynamically counteract heterogeneity interference at multiple hierarchical levels. Extensive experiments on six real-world datasets demonstrate our spatio-temporal prediction method effectively counters heterogeneity, boosting accuracy and robustness. It reduces prediction errors by 4. 53% on average. The method excels in interpretability, efficiency, generalization, and robustness, adapting to diverse patterns and conditions like data sparsity. Its lightweight design ensures real-time efficiency, while interpretable prompts show dynamic adjustments based on learned patterns. This versatility overcomes limitations of traditional methods in handling real-world variations, enabling more reliable predictions for applications from smart cities to climate modeling.