Validación del modelo numérico

3.6

Summary and Discussion

This chapter presents a generic customer segmentation framework that can be used to clas- sify smart meter data into clusters using multiple distinguishing characteristics such as time of consumption, levels of consumption, associated contexts, etc. We also present a clustering consistency index, which can be used to track evolving consumption behaviors and to compare customer segments resulting from different temporal aggregations. Moreover, it is also similar to Rand index [Rand,1971]. However, we generalize it further to determine whether an individual is likely to change cluster over time, i.e., individual to cluster consistency measure.

We evaluate the framework and index using real world smart meter data and survey results. Experiments show that customer segmentation results are different from one context to another. Moreover, different temporal aggregations have only a little effect on segmentation by absolute consumption. But, this does not hold for segmentation by consumption variability or trends. Furthermore, customer’s floor area is relevant to her consumption. In addition, big appliances’ usage patterns also play a role in customer’s consumption variability.

Forecasting Residential Demand

4

One of the important tasks in various Smart Grid applications, from demand response to emergency management, is the short-term electricity load forecasting at different scales, from an individual household to a whole portfolio of customers. While a large body of the load forecasting literature has focused on large, industrial, or national demand, this chapter focuses on energy consumption of residential customers. More specifically, we quantitatively evaluate different machine learning methods for short- term (1 and 24 hour ahead) electricity load forecasting at the individual and aggregate level. Additionally, since energy consumption behavior may vary between households, we first build a feature universe, and then apply Correlation-based Feature Selection to select features relevant to each household. We also find that the improvement provided by the Cluster-based Aggregate Forecasting strategy depends not only on the number of clusters, but more importantly on the size of the customer base. Consequently, our finding provides a valuable insight to practitioners who wish to implement the strategy in the real world.

This chapter revises and extends [Humeau et al.,2013]. The preliminary version of this chapter can also be seen in [Wijaya et al.,2014b].

4.1

Introduction

The exploitation of renewable energy, the integration of distributed energy resources at the distribution level, and the electrification of private transportation are considered as suitable governmental policies to tackle some of the problems of advanced societies, such as reducing CO2 emissions or increasing energy efficiency [Gellings, 2009]. In recent years, these solution

concepts started to pose new challenges to the existing power grids, whose hierarchical, centrally- controlled structure has remained unchanged for a century. For example, the exploitation of renewable sources such as solar or wind may be problematic due to their variable and intermittent nature, while the integration of distributed energy resources may cause congestion and atypical power flows that threaten system’s reliability [Mohd et al.,2008].

In this context, energy consumption prediction for different time horizons (e.g., 1 hour ahead, 1 day ahead, 1 month ahead) and space scales (e.g., distribution transformer, individual house-

50 Forecasting Residential Demand level meter) is becoming crucial for many applications, such as frequency and voltage regulation, demand response (to estimate customer’s baseline [Wijaya et al.,2014d]), and autonomous emer- gency management [Moslehi and Kumar, 2010]. While long-term load forecasting (1-10 years ahead) is important for planning both, transmission and distribution networks, short-term load forecasting (hours to days ahead) is important for the demand response, online scheduling, and security functions of an energy management system. In this chapter, we use the terms energy consumption (or demand ) and load interchangeably.

Overview of Contributions Since energy consumption behavior might vary among house- holds, a feature that are relevant for one house might not be relevant for others. Additionally, we have a large number of houses. Thus, feature selection has to be done automatically. To this end, we first build a (large) feature universe, and then automatically determine the relevant features for each house using the Correlation-based Feature Selection [Hall,1999], which selects subset of features set that are highly correlated with the response variable while having low inter-correlation between each other (see Section 4.3.1).

As we have described in Section 2.1.2, many techniques for energy consumption prediction have been inspired by research on statistical and machine learning, from Linear Regression [Hong,

2010,Papalexopoulos and Hesterberg,1990], ARMA [Huang and Shih,2003,Taylor,2010], and

Generalized Additive Models [Ba et al.,2012,Fan and Hyndman,2012] to Neural Networks [An

et al.,2013,Hippert et al.,2001,Khotanzad et al.,1997] and Support Vector Regression [Chen

et al., 2004, Sapankevych and Sankar, 2009]. However, these techniques have been typically

used at very large space scales, such as predicting the electrical load of a market segment serving thousands of customers or even an entire country. In this chapter, we show that these algorithms can also be used to forecast households’ consumption and improve the benchmark by around 20%–24% (see Section 4.3.2 and Section 4.4). Looking at prediction results, however, load forecasting at the household level remains a hard problem.

Misiti et al. [2010] studied the effect of forecasting clusters of industrial customers to predict their aggregate demand using wavelet-based clustering.1 _{Alzate and Sinn [2013] used kernel}

spectral clustering and consider a mix of residential customers and small/medium enterprises. Interestingly, although Misiti et al., Alzate and Sinn, and our work focus on different customer bases and use different forecasting and clustering algorithms, all conclude that clustering cus- tomers and then forecasting each cluster separately could indeed improve aggregate forecasts. Additionally, our clustering objective is clear, targeting a specific property of the resulting clus- ter (see Section 4.5.1), and we continue by investigating how the improvement provided by this Cluster-based Aggregate Forecasting strategy depends not only on the number of clusters, but also on the size of the customer base. That is, the larger the customer base, the higher the improvement (see Section 4.5.2). Thus, our finding offers additional insight to the practitioners who wish to implement this strategy in the real world.