Low-Frequency Energy Disaggregation Using the Seq2point Method Based on CNN-LSTM Networks

The integration of deep learning into Non-Intrusive Load Monitoring (NILM) technologies has spurred significant advancements in the precision of load identification and power disaggregation. Despite these gains, several hurdles remain, such as low accuracy in decomposing multi-state electrical loads, suboptimal generalizability of algorithms, and lengthy processing times. To tackle these challenges, we present a seq2point load decomposition model that combines convolutional neural networks (CNNs) and long short-term memory (LSTM) networks. This model traverses sequences of total power data to pinpoint the power consumption of target devices within specific intervals of the power data. Our approach begins by employing a sliding window-based event detection mechanism to isolate active load instances, which then serve as inputs for the seq2point model. Using CNNs, the model autonomously extracts local features from the aggregated power data. Additionally, LSTMs are utilized to analyze temporal correlations surrounding each point in the sequence, improving the accuracy of load identification. To evaluate the effectiveness of our model, we applied it to a real-world domestic energy dataset, UKDALE, and compared its performance against leading models. The results demonstrated a substantial improvement of 18.4% in overall performance, confirming the significant enhancement in load monitoring capabilities achieved by our deep learning approach. These findings underscore the potential of combining CNNs and LSTMs to effectively manage the complexities involved in load decomposition, especially in scenarios characterized by intricate and fluctuating power consumption patterns.