<p>Accurately predicting resource load in cloud computing environments constitutes a fundamental challenge for dynamic resource allocation. Traditional threshold-based static scheduling strategies and linear time-series prediction methods struggle to address the nonlinear, abrupt changes and multi-time scale characteristics inherent in cloud workloads. Furthermore, existing deep learning approaches exhibit limitations in terms of noise robustness and multi-scale feature modeling. To overcome these challenges, this study introduces a novel Contrastive Learning Long Short-Term Memory (LSTM) Network model, termed SWT-CLSTM, which integrates Savitzky-Golay (SG) filtering with Smooth Wavelet Transform (SWT). This approach employs SG filters to preprocess and attenuate high-frequency noise, and utilizes SWT for the multi-resolution decomposition of low-frequency trends and high-frequency fluctuations. Additionally, the model incorporates a dual-path neural network architecture, comprising a one-dimensional Convolutional Neural Network (CNN) and an attention-enhanced LSTM. This architecture is designed to extract local patterns and model long-term dependencies. Moreover, the introduction of a frequency-aware hierarchical contrastive learning framework significantly enhances the model’s generalization capabilities for non-stationary data. Experimental evaluations conducted on public cloud task datasets confirm that the SWT-CLSTM model outperforms traditional methods and prevailing deep learning models across various time granularities, thereby markedly enhancing the temporal prediction accuracy of cloud computing resource scheduling.</p>