Household electricity demand forecast based on context information and user daily schedule analysis from meter data

Authors	Yu-Hsiang Hsiao
Title	Household electricity demand forecast based on context information and user daily schedule analysis from meter data
Publication	IEEE Transactions on Industrial Informatics
Volume	11
Issue	1
Pages	33-43
Year	2014
DOI	https://doi.org/10.1109/TII.2014.2363584

Introduction

Background

• Efficient energy usage is an important issue due to the limited energy sources and the increased environmental awareness.
• Electricity load forecasting is a significant factor in achieving economic, reliable, and secure power systems.

Previous Research

• Load forecasting can be categorized into four types, including VSTLF (Very Short-Term Load Forecasting), STLF (Short-Term Load Forecasting), MTLF (Mid-Term Load Forecasting), and LTLF(Long-Term Load Forecasting).
• Many statistical model and AI (Artificial Intelligence)-based models have been proposed to solve load forecasting problems.
• The development of STLF methods is more extensive than VSTLF methods.
• In individual household applications, electricity demand prediction is more challenging because of the lack of an error offset, and context features such as weather, events, economics, and day type are effective predictors of electricity consumption behavior.

Proposed Model

• This paper proposes a novel approach to model the load of an individual household based on context information and daily schedule patterns.

Significance

• The proposed approach can help in understanding an individual household's electricity consumption behavior and precisely predicting the electricity demand.
• This is important for personalized electricity contracts and rates, smart grid and automated demand response applications, and effective electricity deployment.

Proposed Model

• The proposed model involves four stages
• Stage 1 is "Meter Data Collection and Preprocess"
  • Collect and preprocess the time series of daily electricity consumption.
• Stage 2 is "Extraction of Daily Electricity Consumption Behavior Patterns and Pattern Type Induction"
  • Smooth the consumption sharp and impulse by applying the MA (Moving Average) method.
  • Granulate the data into finite granules with discrete intervals.
  • Group similar daily schedule pattern types (or electricity consumption behavior pattern types) into a cluster using clustering techniques.
  • Collect day-dependent context features and investigate features causing electricity consumption behavior to exhibit similar or different patterns.
  • Construct a classification model using only significant context features to predict the likely pattern of electricity demand.
• Stage 3 is "Forecasting Model Construction"
  • Construct intercluster classification model and intracluster prediction model to be used in the proposed procedure
  • Classify the day into one of the consumption pattern types using the estimated day-dependent context features to forecast the electricity demand at a specific time point in a particular day.
  • Use the corresponding prediction model with estimated minute-dependent context features and historical load data as inputs to predict the electricity demand volume

• Stage 4 is "Conducting a Confirmation Test"
  • Intercluster behavior classification model is applied, which uses estimated day- dependent context features to identify the behavior pattern type of the specific day.
  • Next, the intracluster consumption volume prediction model of the behavior pattern type identified is applied, using estimated minute-dependent context features and historical load data corresponding to the specific time point to predict the electricity consumption volume.
  • Finally, a performance index is employed to evaluate the load forecasting error. If the results are unsatisfactory, the adequacy of the context features and constructed models in Stage 3 are checked, while if the results are satisfactory, forecasting applications are initiated.

Experiment

Experiment and Comparison Results

• The proposed approach achieved lower average MAPEs (Mean Absolute Percentage Errors) and lower average MASEs (Mean Absolute Scaled Errors) compared to other methods, indicating that considering household behaviors is a helpful strategy for load forecasting.
• The results of BPN (BackPropagation Neural Network), LR (Linear Regression), and SVR (Support Vector Regression) showed lower forecasting accuracy than the proposed approach, likely due to their lack of emphasis on behavior pattern analysis.
• The SVR2, which means the SVR based on similar historical day, method performed better than SVR by constructing a forecasting model using the pattern-base formed by historical days characterized by high similar context features with the predicted day.
• However, the proposed approach still outperformed SVR2 in terms of average MAPE and MASE.
• The one-step RW (Random Walk) performed the worst among all compared methods, likely due to its heavy reliance on the most recent historical load datum for forecasting.