Action prompt integration in large language models for time series forecasting of nuclear power industry systems

In the main control room of process manufacturing, operators are responsible for monitoring all critical parameters—a setting where the consequences of human errors are unacceptable, especially in high-risk environments such as nuclear power plants. By forecasting post-action trends to proactively assess the correctness of their actions, operators can promptly identify and mitigate potential errors, thereby preventing undesirable outcomes. Most data-driven forecasting methods overlook external factors that may induce abrupt or discontinuous changes. While some approaches attempt to predict by learning action-reaction mappings, they are constrained by limited knowledge bases and poor generalization. Recent advancements in large language models have catalyzed significant transformations in time series forecasting. Accordingly, this paper proposes a novel approach, termed action prompt integration in large language models (APiLLM), for time series forecasting. This approach integrates two inductive biases to enhance the time-series prediction capability of large language models. First, each variable is decomposed into trend, seasonal, and residual components using a local moving-average module trained under the supervision of a global STL decomposition. This explicit separation alleviates the difficulty of the attention mechanism in decoupling mixed-frequency structures. Second, a trainable prompt pool, implemented as a key-value memory, retrieves relevant prompt embeddings based on the learned summary representations of individual patches. These prompts are prepended to the component tokens, enabling the model to reuse transferable temporal patterns and thus improve generalization. Additionally, the attention module is augmented with an action-aware additive bias that converts action-time encodings into a structured adjustment of the attention scores. This selectively enhances attention around actions while remaining fully compatible with standard scaled dot-product attention (SDPA) backends. Finally, in the Tennessee Eastman process simulation, human actions are emulated through closed-loop control of setpoints, whereas external disturbances in the nuclear power circulating water system are modeled as abnormal operating events, including outlet pipeline leakage and packing cooling anomalies. Experimental results show that APiLLM can accurately predict post-action parameter trends, thereby enabling action correction and supporting operational planning.