Clean, high-resolution time series data is essential for precise predictions and effective control in smart building energy systems. Based on a refined WaveNet, this study proposes a novel architecture for unsupervised anomaly identification in univariate energy consumption time series. Without labeled data, the model is specifically intended to detect abnormal patterns and capture long-range temporal correlations. It uses gated activations, skip connections, and dilated causal convolutions to improve temporal fidelity, resilience, and sensitivity. To the best of our knowledge, this is the first anomaly detection strategy in this setting that integrates architectural improvements with a thorough assessment of the impact on downstream forecasting. High-frequency energy consumption time series are used to rigorously benchmark the Refined WaveNet against cutting-edge baselines, such as the original WaveNet, deep generative hierarchical learning (DGHL), and variational autoencoder (VAE). With a Precision-Recall Area Under the Curve of 99.23% and an F1 Score of 98.30%, the model outperforms the original WaveNet—which achieves only 41.96% in F1 Score—by more than 56 percentage points. We combine the detection module with a long short-term memory (LSTM) forecasting model in order to evaluate its practical usefulness. According to experimental findings, adding just 10% synthetic anomalies to the time series increases mean squared error (MSE) by more than thirty times, whereas using Refined WaveNet for preprocessing brings forecasting performance back to levels that are almost pristine. These results highlight anomaly detection’s significance as a fundamental element of time series analysis for intelligent energy systems and establish Refined WaveNet as a small, effective, and deployable solution for edge-based, real-time applications. The code is available at: https://surl.li/oumfna.
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号