Non-intrusive monitoring of electrical loads based on switching transient voltage analysis: Signal acquisition and features extraction

Non-Intrusive Load Monitoring (NILM) based on switching transient voltages was introduced in the last decade as part of an initiative to develop easy-to-install NILM systems with minimum number of sensors. The voltage transients induced by the connection or disconnection of different appliances propagate through the electrical wiring and can be measured at one point in the house. These transients can be used to identify the appliance in operation and the state of its switch. The existing solution makes use of a single Notch analog filter to remove the 60 Hz component before the voltage signal is digitally acquired. The transient is detected through the variation in the Short Time Fourier Transform (STFT) continuously computed on 1microsecond adjacent time windows. Thereafter, the duration of the transient and the averaged magnitude of the STFT form a feature vector that is classified using a Support Vector Machine (SVM) model. This dissertation proposes to improve this solution by reducing the computation burden to detect the transients, by including multi-Notch digital filtering to remove not only the 60 Hz component but also the harmonic interference, and by using Wavelet transform methods to compute a reduced feature vector to classify the switching transients. The Envelope or Mask Trigger detection method is examined in this dissertation and presented as a suitable trade-off between hardware complexity, computation burden and detection accuracy. Once the transients are detected and recorded, a cascade of discrete IIR second order Notch filters is designed to remove the power harmonic interference. This filtering is highly affected by the natural response of the filters. The state of the art methods to suppress the natural response are effective but at the expense of considerable computation time or degradation of the frequency response. A novel natural response suppression method based on projecting the zero state response is proposed in this dissertation. This method is demonstrated to overcome the drawbacks of the previous techniques on both simulated and measured transients. The filtering of the power harmonic interference is addressed by a twofold solution. The first part is to use a cascade of independently designed second order Notch filters. The exclusive dependence of the magnitude of the poles on the Notch bandwidth is stated and used to derive boundaries for the bandwidth and the natural response duration specifications to design second order Notch filters. On the other hand, if the degradation of the cascade frequency response is not tolerable, an Infinite Impulse Response (IIR) multi-Notch filter design is recommended. The existing design methods does not guarantee unity transmission between Notch frequencies along with exact compliance with either bandwidth or cut-off frequency specifications. In this dissertation, a new flexible method to design IIR multi-Notch filters based on polynomial pole placement is demonstrated. The filter is designed by determining a shape factor and the local maxima positions of the frequency response magnitude. The shape factor can be computed to comply with either cut-off frequencies or bandwidths specifications at any gain. The designed filter is stable, ripple-free, and can be implemented using an all-pass filter. The optimal selection of the shape factor and the local maxima positions by using Particle Swarm Optimization (PSO) is demonstrated as well. Finally, the use of Continuous Wavelet Transform (CWT) and Wavelet Packet Transform (WPT) in order to classify, via SVMs, transient voltages is discussed and compared with the previous approach based on STFT. The influence on the classification accuracy of the mother wavelet, the selection of scales or decomposition levels, and the time integration or aggregation methods is discussed and experimentally tested. Two sets of heuristic guidelines to extract features from switching transients by using either CWT or WPT are derived. In the experiments, the CWT and the WPT methods achieve comparable maximum classification accuracies and both outperform the previous methods with considerably shorter feature vectors.