A CNN–LSTM Architecture with Residual and Squeeze-and-Excitation Blocks for Scenario-Based Non-Intrusive Load Identification

Received: 7 October 2025 | Revised: 12 January 2026 | Accepted: 31 January 2026 | Online: 4 April 2026

Corresponding author: Aekkarat Suksukont

Abstract

Non-Intrusive Load Identification (NILI) is a crucial technique in energy management, enabling the reduction of unnecessary energy consumption and supporting the development of smart building systems. Nevertheless, accurately distinguishing electrical appliances with similar operational characteristics and addressing the complexity of aggregated electrical signals are challenging tasks. This study proposes a deep learning-based NILI framework designed to effectively extract discriminative features from energy consumption data. The proposed Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) architecture incorporates Residual Blocks (RB) and Squeeze-and-Excitation (SE) blocks to enhance feature representation while alleviating information loss during deep feature propagation. The network comprises three convolutional blocks with integrated SE layers to strengthen channel-wise feature attention, followed by an LSTM module that captures long-term temporal dependencies in sequential energy signals. Unlike conventional appliance-level disaggregation approaches, the proposed framework performs scenario-based classification, where each class represents a unique combination of simultaneously operating electrical appliances. The model is trained and evaluated on datasets consisting of operational scenarios involving two, three, and four electrical devices, with raw electrical signals transformed into kurtogram representations to emphasize salient signal characteristics. The experimental results demonstrate that a learning rate of 10⁻³ consistently outperforms 10⁻⁴, achieving training accuracies of 98.04%, 99.95%, and 99.75%, along with precision values of 99.98%, 95.65%, and 97.10% for the respective scenarios. In contrast, the lower learning rate leads to noticeable performance degradation and higher residual training loss. These results confirm the robustness and effectiveness of the proposed NILI framework and highlight its potential for practical deployment in intelligent and sustainable energy management systems for future smart home environments.