A Short-term Load Forecasting Method for Amusement Parks Based on K-means Clustering, Holiday Features and Long Short-Term Memory (LSTM)

Amusement parks feature complex power consumption structures, and their passenger flow is highly susceptible to disturbances from weather and holidays. Consequently, their power load exhibits characteristics including strong randomness, prominent intraday periodicity, distinct spike mutations, and mixed load patterns, which makes it difficult for traditional single prediction models to accurately capture its fluctuation patterns. To improve the accuracy of short-term power load forecasting for amusement parks, this paper proposes a load forecasting method that integrates K-means clustering and multi-feature Long Short-Term Memory (LSTM). First, this paper analyzes the temporal characteristics and forecasting difficulties of amusement park load, and constructs a multi-dimensional feature system covering time, meteorological, and holiday classification features. A four-level classification rule is adopted to realize fine-grained holiday grading, which addresses the problem of insufficient granularity of the traditional binary classification. Then, the unsupervised K-means clustering algorithm is introduced, combined with the silhouette coefficient to select the optimal number of clusters, enabling the automatic partitioning of daily power consumption patterns. Finally, an LSTM temporal forecasting model is constructed based on a sliding window with a window size of 24, to achieve hour-level dynamic temporal inference. Case study results show that the proposed optimized model yields a Symmetric Mean Absolute Percentage Error (SMAPE) of 8.06%. Compared with the standard single LSTM model, this method effectively mitigates issues including mixed load patterns, peak load underestimation, and over-smoothing of forecasting curves, with significantly improved fitting accuracy and generalization capability.