STLLM-GAN: Spatio-temporal LLM Generative Adversarial Network for PM2.5 prediction

Accurate $P{M}_{2.5}$ prediction plays an important role in climate change mitigation and environmental protection. As an emerging artificial intelligence technique, Large Language Models (LLMs) have exhibited powerful data processing and adaptive feature learning capabilities, thus being widely applied in Time Series Prediction (TSP). Unfortunately, the current LLM-based TSP models pose difficulties in accurate $P{M}_{2.5}$ prediction due to two factors: 1) they neglect or fail to fully extract spatio-temporal dependencies, and 2) these LLM-based TSP models solely rely on the supervised training, failing to learn real data’s distribution. Jointly considering these factors, in this paper, we present a novel $P{M}_{2.5}$ prediction framework namely Spatio-Temporal LLM Generative Adversarial Network (STLLM-GAN). In detail, to capture spatio-temporal dependencies, Spatio-Temporal Large Language Model (STLLM) is first developed, containing a Spatio-Temporal Module (STM) and an LLM-enabled Inference Module (LLMIM). For the purpose of optimizing STLLM’s training, we design an adversarial training scheme using Generative Adversarial Network. The scheme incorporates un- and supervised training through a jointly beneficial manner. The unsupervised training seeks to learn real data’s distribution, while the supervised training strives to align the actual values with estimations by minimizing Mean Squared Error (MSE) function. We carry out extensive experiments on two real-world air pollutant concentration datasets, covering Shanghai city and Beijing city, respectively. The experimental results prove that STLLM-GAN is superior to advanced benchmarks in prediction performance.