Accurate non-intrusive load monitoring forecasting using high frequency sampling and eventless decomposition approach

Electricity is now the primary energy source for humans, surpassing transportation and industry. Lowering a building's energy usage is essential for enhancing performance. Load forecasting (LF) aids in achieving energy efficiency (EE) goals such as reducing overall load and peak load clipping. Direct monitoring of all electrical equipment is essential for implementing these objectives. Direct monitoring is expensive and unsustainable due to the numerous sensors needed for each appliance, posing a significant obstacle to LF implementation. An alternative method, called non-intrusive load monitoring (NILM), has been proposed. NILM can monitor multiple equipment simultaneously by analyzing data from a single meter's total load readings. This method utilizes data from a single meter's total load readings to monitor multiple equipment simultaneously. The project employs MATLAB software, including Classification Learner, Data Cleaner, and Neural Networks Toolboxes. The results showed the model can disaggregate and forecast the power consumption of the non linear loads. Firstly, harmonics can generate different values even if the loads are similar in consumption and outperform the other features. Compared to Bagging Decision Trees and Gaussian Naive Bayes, K-Nearest Neighbors (KNN) classifier performed better. KNN classifier had 98% phase A, 88.8% phase B, and 99.8% phase C accuracy which is the suitable technique of electrical load disaggregation for NILM. However, the model requires further enhancement as the classifier encounters accurately disaggregate similar load values. To address this, a moving median filter (MM) was introduced to reduce noises and minimize errors. Additionally, a nonlinear autoregressive network with exogenous inputs (NARX) model was employed to forecast each load 24 hours in advance. The results of root mean square error (RMSE) before the MM filter represented 0.471, 0.5291, and 0.6269 based on phase A, B, and C respectively. Moreover, the findings of RMSE after the MM filter accounted for 0.089, 0.0904, and 0.4484 based on phase A, phase B, and phase C respectively. The reason behind the high error based on phase C is the noise generated by the NILM and the difficulty of smoothing it enough. NARX excels at forecasting non-linear signals with high accuracy. The NILM system findings were utilized in forecasting to see how the technique selected for NILM affects the forecasting accuracy, both before and after applying the MM filter. The addition of the MM filter reduced the error rate. Furthermore, integrating NILM with LF resulted in successful load forecasting.