Non Intrusive Load Monitoring

Transcription

1 Non Intrusive Load Monitoring Felice Tuosto

2 Non-Intrusive Load Monitoring (NILM) Disaggregation of individual appliances from the aggregated energy consumption data collected by a smart power meter. Appliances by energy consumed (US): difficulty to disambiguate from noisy signal. Source

3 Objective Carry out HW/SW product to estimate real time and offline electrical consumption of known devices starting from aggregated loading signals Know in real time what energy you are using and how much it is costing you.

4 Introduction Different devices yeld different types of signals: On/Of On/Off-Multi-States On/Off-Growth/Decay On/Off-Multi-States-Growth/Decay Variable-Consumption Composite-Model

5 Methodology Event-based methods try to detect On/Off transitions Non event-based methods try to detect whether an appliance is On during the sampled duration...

6 Methodology finding the best solution Algorithm Computational Complexity Memory Requirements FHMM O (K2N) K2N CFHMM O (K2N) K2N HSMM O (KN) KN. D ANN & SVM O (K3N) K3N CO O (KN) KN

7 Proposed solution: model flow RealTime Reader Denoising Filter KF State Detection Is Steady State Yes Steady-State Signature Extraction No Transient Signature Extraction Transient Coherence Model Steady-State Coherence Model(*) Energy Disaggregation Signature DB Signature DB Correction (*) Delete the combinations of states that may not correspond to the current aggregated load => It reduces the search area of the solution (to only feasible combinations)

8 KF State Detection Algorithm Steady-State: operative state of an electrical device with an almost-constant consumption Transient-State: else Tasks: Kalman Filter: mean(p), Var(P) Var(P)>Th => transient state

9 Methodology Off-line Phase Real-time Phase Appliance Signature Identification Algorithm Real Time Disaggregation Algorithm de-noising filtering state detection (KF) clustering of devices (GMM) pdf estimation signature database writing Aggregated time series segments are decomposed into a combination of time series segments (maximizing the a posteriori probability to observe that combination). signature database reading

10 Appliance Signature Identification Algorithm Consumption Signature: Unique Features of the devices A parameter computable through the analysis of the disaggregated consumption. transient signatures Describing the behaviours of the electrical transitories steady-state signatures Describing the behaviours of the electrical system states and the variation between different states A set of signatures describes univocally a device!

11 Appliance Signature Identification Algorithm: GMM 1 device K steady states K gaussians K clusters Raw Signal works on disaggregated signal time series available in training phase (test +open data) probabilistic model: data generated by a combination of Gaussian random variables => each device is a mixture of k gaussians: Filtered Signal k [ k, k, wk ] [... ] k1 kn f k [ 1... K ] Diagonal covariance matrix ( ) => v.a. independent 2 2 k [ k 1... kn f ] w [ w...w ] k1 kn f k Nf p x k p x ki i 1 GMM θ Training GMM: Expectation-Maximization (EM) computes the parameters: Computational costs = O(K) Features of the devices

12 Appliance Signature Identification Algorithm: GMM 1 device K steady states K clusters K gaussians Through the Expectation Maximization algorithm, k clusters are identified and expressed in terms of arrays of means and variances. One device: 2 states state 1 state 2 Many devices: many states

13 Learning with Python

14 Appliance Signature Database Off-line Phase table of all means and variances for each feature, e.g. active power, reactive power, etc... each row is a state of the device Signature Database update example: APPLIANCE NAME ACTIVE POWER MEAN ACTIVE POWER STD REACTIVE POWER MEAN REACTIVE POWER STD 1 harmonic 3 harmonic Dremel VacuumCleaner Refrigerator Refrigerator Phon

15 Real Time Disaggregation Algorithm Real-time Phase Raw Signal performed at each sample period provides the signal disaggregation Filtered Signal Identifying the best combination of the state of devices registered in the signature database. Signature DB Signal Disaggregated COMPUTATIONAL COMPLEXITY O (KN) A11 k-th combination of gaussian: A12... Source: Aggregated Procedure of Gaussian Mixture Models for additive features Ck Air A js [0,..1,0,..1,0..] Air ck A... A js Ck N ( Aik A jh, A2ik A2 jh )...

16 Real Time Disaggregation Algorithm The best combination of gaussian variables is found maximizing: max p Ck x(t ) k p C k p x ( t ) C k i p Ci p x ( t ) Ci p x (t ) Ck N x(t ); Ck, Ck Copt := best combination (mixture of occurrences of the devices)

17 Learning with Python

18 Energy Disaggregation Consumption APPLIANCE NAME TIMESTAMP START TIMESTAMP STOP ESTIMATED ACTIVE POWER [Wh] Dremel :00: :45: VacuumCleaner :00: :45: Refrigerator :00: :45: Refrigerator :00: :45: Phon :00: :45:

19 Experimental results Output: From the aggregated signal, algorithms identify any single device

20 Next Steps Additional features to the existing models (wavelets,) Advanced methods (Deep Neural Networks) with: One network for each appliance Real aggregated data + synthetic aggregated data (GMM) window