Modelling Framework for NILM Bo LIU 1 , Wenpeng LUAN 1,2 , Yixin YU 1 - - PowerPoint PPT Presentation

A Fully Unsupervised Appliance Modelling Framework for NILM

Bo LIU1, Wenpeng LUAN1,2, Yixin YU1 1. School of Electrical & Automation Engineering, Tianjin University, Tianjin, China 2. China Electric Power Research Institute, Beijing, China 3rd International Workshop on Non-Intrusive Load Monitoring May 15 2016 in Burnaby , BC, Canada Tianjin University

Tianjin University

NILM

 NILM analyses the aggregate electricity usage data measured at the power supply entrance of the electric load to acquire the appliance- level specific consumption information via pattern recognition techniques and machine learning methods.

Source: G.W. Hart “Nonintrusive appliance load monitoring” (1992)

Tianjin University

NILM

Origins from G.W. Hart “Nonintrusive appliance load monitoring” (1992)

 Principles

① ② ③ ④

Load Signature Extraction Appliance Modelling Load Behavior Monitoring Appliance Naming Appliance Signature DB ① ③ ② ④

Tianjin University

Introduction

 Load Modelling is the prerequisite for implementing NILM  Finite state machine(FSM) is the most common adopted method

Source: G.W. Hart “Nonintrusive appliance load monitoring” (1992)

 Supervised Modelling (supervised parameter learning from a complete set

• f labeled signature samples)

 Semi-supervised Modelling (unsupervised parameter learning with the knowledge

• f the state set and topology of the model)

 Fully unsupervised Modelling

Tianjin University

Introduction

 Fully unsupervised appliance modelling

Get complete state set, topological structure and model parameters of FSM without any priori knowledge from the aggregate load data

Tianjin University

Framework Overview

 Fact and Assumption: repeated appliance behavior patterns  Ideas:

Simple Cycle Event Sequence(SCES) pattern constructing FSM Use the frequent pattern mining techniques to extract appliance SCES-patterns and combine them

Figure drew by the power data from the public dataset “Blued”

FSM Model Topology of a three-state hair drier

（停机）

2

（加热2）

1

（加热1）

• ff

Low-grade Heating High-grade Heating

Samples/S(fs=1Hz) Power/W

refrigerator

Tianjin University

 Five Steps：

Detection of IsoES Clustering and Labelling of load events Mining the SCES patterns Grouping of the SCES patterns FSM model topology generation and parameter estimation the set of ΩIsoES logical name of event the set of ΩSCES groups of SCES patterns

Framework Overview

Tianjin University

Proposed Methods

Step1: Detection of IsoES：

 Isolated load Event Sequence (IsoES)  IsoES detection method  Taking IsoES as event sequence record in ESDB

Tianjin University

Proposed Methods

Step2: Clustering and Labelling of load events(in ESDB)

Get a unique logical name for each event  mean-shifting clustering method any clustering analysis method not requiring cluster number  signature vectors representing event e

< (e11,t11), (e12,t12), (e13,t13), (e14,t14), (e15,t15) > < (6,t11), (1,t12), (3,t13), (2,t14), (5,t15) >

Unlabeled IsoES Event-labeled IsoES

Tianjin University

Proposed Methods

Step3: Mining the SCES patterns

Frequent Event Sequence (FES) patterns mining  Class GSP (Generalized Sequential Pattern) algorithm Filtering out the SCES patterns with the β-ZLSC constraint

Tianjin University

Step4: Grouping of the mined SCES patterns

Divide the acquired SCES patterns into different groups associated with different appliances

 Fact and Assumption:

the load events produced by different appliances are different

Event Correlation Rule

Proposed Methods

Tianjin University

Step5: FSM model topology generation and parameter estimation

Incremental Topology Generation and Parameter Estimation (ITGPE) 12 3 5 3 10 5 3 10 12 < 12,3 > < 10,5,3 >

Proposed Methods

TG PE

Tianjin University

Experiments

76 IsoESs

Aggregate load data (24h, 1Hz)

12 3 5 3 10 5 3 10 12 < 12,3 > < 10,5,3 >

Tianjin University

Conclusions

 A fully unsupervised appliance FSM modelling framework is proposed and validated on real measured data  The applicability of the existing NILM technologies is improved  Moving towards the realization of the auto-setup NILM

Tianjin University

Future work

 Comprehensive testing and analysis on more measured data  Improvement on the methods and algorithms used by different modules of the framework  Appliance Naming for the acquired FSM model

Q&A

luanwenpeng@epri.sgcc.com.cn

Tianjin University