两阶段深度学习、随机森林与回声状态网络非侵入式负荷分解研究

Two-Stage Deep Learning, Random Forest, and Echo State Network for Non-Intrusive Load Decomposition: A Comparative Study

摘要: 在全球碳中和发展趋势下，新型电力系统对需求侧用电精细化管理提出了更高要求，非侵入式负荷分解(Non-Intrusive Load Monitoring, NILM)作为用户用电行为感知的核心技术，成为智能电网领域的研究热点。针对当前NILM领域研究多聚焦于单一模型的结构优化，缺乏不同技术范式模型在统一基准下的系统性横向对比，工程应用技术选型缺乏量化参考的问题，本文对两阶段深度学习、随机森林、回声状态网络(Echo State Network, ESN)和隐马尔可夫模型(Hidden Markov Model, HMM)四类分别代表分层深度学习、传统集成学习、轻量化储备池计算和传统概率图模型范式的NILM方法开展系统性研究。首先，构建了NILM任务的数学模型与物理约束体系，针对四类模型的核心特性，分别完成了适配负荷分解任务的结构设计、特征构建与参数配置；随后，在统一原始数据、统一可比样本、统一评价指标体系和统一后处理规则下，从分解精度、负荷适应性、抗噪声鲁棒性、推理效率与轻量化性能四个维度开展对比实验。实验结果表明，随机森林模型在无噪声场景下分解精度最优，平均MAE为0.2064、平均R²达到0.9573；两阶段深度学习模型综合性能次之，平均MAE为0.6101，并在5%输入噪声下仍保持较小性能波动，表现出较好的鲁棒性；HMM模型整体精度弱于前两者，但具有参数量少、推理速度快、可解释性强等优势，可作为传统NILM方法的代表性基线；ESN模型具备训练机制简单和轻量化潜力，但在当前实验设置下平均MAE为2.4199、平均R²为−5.4685，基础分解精度不足，难以满足实际工程应用需求。本文研究明确了四类模型的优劣势、适用边界及工程定位，可为不同资源约束和不同精度需求下的NILM技术选型提供系统参考。

Abstract: Against the backdrop of global carbon neutrality, the emerging power system has imposed higher requirements on the refined management of demand-side electricity consumption. As a core technology for perceiving user-side electricity-use behavior, Non-Intrusive Load Monitoring (NILM) has become a major research focus in smart grids. To address the fact that most existing NILM studies concentrate on optimizing a single model, while lacking systematic horizontal comparisons among different technical paradigms under a unified benchmark, this paper conducts a comparative study of four NILM methods: two-stage deep learning, Random Forest, Echo State Network (ESN), and Hidden Markov Model (HMM), representing hierarchical deep learning, traditional ensemble learning, lightweight reservoir computing, and conventional probabilistic graphical modeling, respectively. First, the mathematical formulation and physical constraint system of the NILM task are established. Then, model design, feature construction, and parameter configuration are carried out according to the characteristics of the four methods. Subsequently, comparative experiments are conducted under a unified raw dataset, unified comparable samples, unified evaluation metrics, and unified post-processing rules, from four perspectives: disaggregation accuracy, load adaptability, noise robustness, and inference efficiency with lightweight performance. The experimental results show that the Random Forest model achieves the best disaggregation accuracy under noise-free conditions, with an average MAE of 0.2064 and an average R² of 0.9573. The two-stage deep learning model ranks second overall, with an average MAE of 0.6101, and maintains only a small performance degradation under 5% input noise, demonstrating strong robustness. The HMM model is less accurate than the above two methods, but it offers clear advantages in terms of interpretability, parameter efficiency, and inference speed, making it a representative traditional NILM baseline. In contrast, the ESN model, although simple in training mechanism and potentially lightweight, yields an average MAE of 2.4199 and an average R² of −5.4685 under the current experimental setting, indicating insufficient basic disaggregation accuracy for practical engineering applications. Overall, this study clarifies the strengths, limitations, application boundaries, and engineering roles of the four methods, and provides a systematic reference for NILM model selection under different resource constraints and accuracy requirements.