Mining volunteered geographic information for predictive energy data analytics

Big data analytics in the energy industry

Utility companies hold, however, millions of data points on their customers that can be developed to a valuable business asset. This available data will further increase in the future due to the roll-out of smart meter and Internet of things infrastructures. Techniques of big data analytics, especially predictive information systems will help utility companies to make sense of that data (Sharma et al. 2014; Zhou et al. 2016) and can enable them to create innovative products and services. Even though, first steps in development of algorithms and tools for predictive energy data analytics are done and basic knowledge on how to create insights from big data exists, but the currently existing models are often not reliable enough to be used in practice. Therefore, further research on better predictor variables, more suitable data processing and knowledge modeling techniques is necessary. This work contributes to this research gap and investigates the use of freely available geographic data for predictive energy data analytics in the specific but relevant example of household classification.

Predictive analytics in energy retail: the use of household classification

A possible reaction to the harsher market requirements for energy retailer is to improve the customer communication and to offer product or services that better fit to the needs of their customers (Gebauer et al. 2014), establish targeted customer communication or one-to-one marketing. Examples for such products are tariffs for specific customer groups, energy consulting, and energy- or infrastructure-related cross-sales (e.g., heating and cooling appliances, photovoltaic installations, Internet access). For the development and commercialization of such products and services, detailed knowledge about customers is necessary. Typical variables needed in marketing campaigns are household characteristics (e.g., families with children, household type and size) or knowledge on the customers’ attitudes and intentions (e.g., towards sustainability, purchase intentions for products). Recent studies have shown, that various household characteristics (such as household type, type of the heating system, number of residents, etc.) can be identified using 30-min electricity smart meter data (Beckel et al. 2014). Based on 15-min smart meter and hourly weather data, household characteristics related to energy-efficiency can be identified (Sodenkamp et al. 2017; Hopf et al. 2018). Other studies show that households with old heating systems can be identified to be addressed in an energy-efficiency campaign (Kozlovskiy et al. 2016), or customers using an electric vehicle (Verma et al. 2015), which is also an interesting insight for energy utilities. In all studies on household classification, several attempts have been made to improve the quality of predictive artifacts. Multiple data preparation and machine learning algorithms have been tested so far, but the models still need further improvements and lowered error rates to become more reliable and ready for real applications in the industry. I argue that additional predictor variables need to be taken into account to increase the quality of household classification models, since the information that is possible to extract from electricity consumption data is limited. In this paper, I therefore present VGI as a promising source of additional data that is freely available online and has emerged in the past years.

Volunteered geographic information as a data source for predictive analytics

A significant number of web portals arose during the past years, due to the fact that users collaborate and create crowd-sourced data that often has a geographic reference associated. Goodchild (2007) coined the term VGI for this phenomenon of user-generated geospatial data. His work attracted a considerable amount of research in the past ten years and lead also to attention in other disciplines than geography. Unique features of this data source are the free availability and the variety of subjects – even for some subjects, data was never-collected before (Sester et al. 2014).

In this work, all user-generated geospatial data is considered as VGI as long as it was collected with the (explicit or implicit) consent of the user. This definition does not include data provided or published in a non-voluntary way (e.g., accidentally published or revealed by hackers).

This definition is necessary here, since there is no clear consensus in related works on the concept of VGI. On the one hand, there is the understanding of VGI in a strict sense, restricting it to user contributions that are made “with the intent of providing information about the geographic world” (Elwood et al. 2012). Examples for this interpretation of VGI are the co-creation of maps (e.g., OpenStreetMap, Wikimapia) or the collection of environmental data (e.g., Christmas Bird Count). On the other hand, user contributions without the intrinsic motivation to publish geographic data exist, since users create for tweets, posts, likes, reviews, photos or other content that has a spatial datum. Stefanidis et al. (2013) name such information Ambient Geospatial Information (AGI), but they are seen as VGI in a broader sense in this paper. Harvey (2013) suggests to delimit the term VGI to data that was collected with a clear opt-in of users. He suggests the term Contributed Geographic Information (CGI) for all geographic data that was collected without that clear consent, but having an opt-out possibility for users as. Harvey names “cell phone tracking and RFID-equipped transport cards” as examples, but the use of such data, outside the objective it was collected for, stays obviously often in conflict with privacy or data protection regulations. Therefore, the above given definition of VGI in contrast to non-voluntary generated data is preferred here.

A vast number of VGI initiatives are already documented in the literature: (Elwood et al. 2012) mention 99 initiatives that they found in 2009 and categorize them according to geographic extend, the date initiated, the sponsoring entity and the primary purpose of the initiative. Ballatore et al. (2013), Sester et al. (2014), such as Rinner and Fast (2015) document further platforms and suggest different categorization schemes for initiatives, by focusing on the usage of VGI data, the data types and formats used. Due the dependence on continuous contributions from users, a large turnover of newly emerging, changing and perishing projects is visible. Based on the existing collections (Elwood et al. 2012; See et al. 2016), I collected 116 VGI initiatives that have been documented in the literature and analyzed them together with students. The list was further extended with a number of real estate advertisement portals where users can search for hiring, renting, selling or buying houses and dwellings. In total, we identified a sample of 127 initiatives. This collection is not intended to be an extensive list of VGI initiatives, but to give some idea of the types and categories of such data sources. From the 116 VGI initiatives that stem from the literature, only 64 (53.7%) still have an active web site or could be found via the Internet¹. However, some initiatives have become very popular and attracted many active users (such as OSM, or Geocaching).

By looking at the distribution of active initiatives according to data types (Rinner and Fast 2015) and categories (Sester et al. 2014) in Fig. 1, one can see that the majority of projects deal with location data, mapping, or environmental and social media topics. The category “business” is also frequent, but this figure is misleading, since ten German-speaking real estate portals were assigned to this category that might over-represent this category in comparison to others. Generally, VGI initiatives contain a considerable amount of data that is related to an energy context and is therefore relevant for the utility industry, given that knowledge from the freely available data sources can be derived. However, the mining of such data sources is associated with large effort to access the data. For example, only 26 of 75 active platforms have an explicit API for data access.

So far, the use of VGI data in the energy domain is sparse, but could bring value to energy retailers and service providers in the utility industry. This paper aims therefore to motivate the use of VGI, illustrates the usage of such data and the possible contribution for predictive data analytics in the field of energy retail considering two prominent examples of VGI initiatives.

Empirical and automatic dimensionality reduction

Model building is one of the core techniques used in big data analytics (Abbasi et al. 2016), since knowledge, insights and business value does not automatically come from simply applying analytic tools to data. It must be leveraged by analysts and managers into strategic and operational decisions to generate value (Müller et al. 2016; Sharma et al. 2014). One of the major challenges in sense-making from big data is to identify the relevant features and encounter the “curse of dimensionality” (Keogh and Mueen 2011) lowering the quality of predictive systems. This problem is serious: Whereas in the last decade, a number of 50-100 features was called a “large” feature set (Kudo and Sklansky 2000), today we are confronted with hundreds or even thousands (Hua et al. 2009) of features.

Much research has been conducted on automatic dimensionality reduction – also known as feature selection. The interested reader is referred to the general introduction of Guyon and Elisseeff (2003) or the comprehensive literature reviews made by Chandrashekar and Sahin (2014), Liu and Motoda (2008) or Saeys et al. (2007). However, the capability of such automatic feature selection algorithms is limited, at least for the investigated class of machine learning problems in energy retail. Empirical definition of relevant features is therefore a cornerstone in the development of such tools, since it enables the modeling of human knowledge into the data. By defining features empirically, also expert knowledge or commonsense can be coded into the data.

Electricity consumption features

Geographic features from OpenStreetMap data

The empirical definition of features for VGI map data is done using the currently largest VGI initiative OSM (Jokar Arsanjani et al. 2015) as the study subject. The data in this project is – typically for geographic map data – characterized by multiple spatial dimensions and contextual data that cannot be easily used in data analytics methods. In detail, the data in OSM consists of points, poly-lines and spatial relations that are annotated with tags (containing plain-text data) to give semantic meaning to the objects. Therefore, feature extraction a necessary step making the data accessible in energy data analytics.

The feature extraction is done using problem specific knowledge applied to the data to derive meaningful variables (Guyon and Elisseeff 2006): Some features are directly derived from the raw data (e.g., frequencies of objects in an area). Besides that, a literature survey in geographic information science was conducted and spatial landscape metrics (e.g., describing the spatial envelope or the structure of a landscape) were identified as domain-specific features. The identified features were grouped into topologic features describing the structure of and relations between one household and spatial neighbors (e.g. longitude, latitude, frequency objects in the surroundings, distance to city center), and semantic features considering the meaning of an object within the spatial context it appears. In the group of semantic features, I differentiate between point-based features and polygon-based features (i.e. building and land use). In the remaining section, the categories of features are described.

Topologic features — In geodesy, topology focuses on the structure of and relations between spatial objects, e.g., whether two objects overlap or whether one object is contained in another (Becker 2012). In contrast to the semantic relation of a geographical object to another, the topology covers only shape and arrangement of objects. Example topologic features are geographic location (longitude, latitude, altitude) that representing the local environment, such as climate conditions or the number of sun hours per day. The number of objects in the surrounding of a household is a proxy for the population density and distance to city center is a distinction criterion for urban or rural area

Semantic features — A charted point or polygon in a map can have different meaning, depending on the size of the object, in which region it appears and which objects are nearby. The semantics focuses on the meaning of an object within the context it appears. The available source for semantic geographical data are tag-annotations associated to objects in OSM.

It is obviously difficult to express the proximity of two or more objects by means of a semantic distance. Suggested similarity measures, as used for example in information retrieval (Schwering 2008; Janowicz et al. 2011; Ballatore et al. 2012), can not be easily adapted in the context of this study, since they mainly express the similarity of two objects, but not the similarity of areas on a map as needed for the characterization of households. Therefore, I defined semantic features from the OSM internal taxonomy for tags of geographic objects that is known as OSM map features (the double meaning of this term cannot be avoided here) that is maintained by the OSM community.

Feature extraction from real estate portals

As a second source of VGI data, publicly available online real estate advertisements are considered. This data is user-generated and has – in most cases – a geographic location associated. In the context of household classification, the use of this data source seems natural, given that household and building characteristics are often similar in neighborhoods.

The major issue regarding the use of real estate portal data is the sparsity of data (i.e., high number of missing values). There are two reasons for missing values in features from that data source. First, the number of real estate advertisements in the surrounding of a target address is strongly dependent on the location (there might be, for example, more advertisments in city centers than in suburbs) and on the point in time when the data is retrieved (e.g., more offers at the end of school years and less offers at the start of courses in university cities). Second, the amount of information provided in advertisments varies to a large extent, because users only insert information on household properties that exist (e.g., elevator in the house, pool in the garden) or do not want to provide some information.

For a sample of 4,521 household locations in Switzerland, 8,341 advertisements from a Swiss real estate portal with 133 different attributes have been downloaded with a web-crawler between March 23 and May 10, 2017. Thereby, all advertisments within a radius of 1,000 m around each household address were considered. Figure 3 illustrates the availability of advertisements for the considered locations. For the majority of advertisements, only basic attributes are present (e.g., living area, rent per month, rent per month, textual description). Table 2 shows the most frequent attributes that are present in more than 50% of the advertisments. Infrequent attributes, only avaliable in less than 1,000 advertisments, were discarded to not distort the analysis and being able to obtain generalizable results in this exploratve study. Future studies may try to impute values based on further data (Saar-Tsechansky and Provost 2007), or infer the information from textual descriptions using text mining.

Attribute	Data type	Missing values (in %)
Description	Text	1.39
Main dwelling category	Categorical (e.g., house, dwelling, business)	2.01
Type of dwelling	Categorical (e.g., studio, single-family home, shared flat)	2.01
Geographic location	Coordinates	22.35
Amount rooms	Number	24.13
Living area	Number	30.64
Floor	Categorical (e.g., 1st, 2nd, attic)	32.77
Rent per month	Number	36.14
Area living flat	Number	42.86
Additional costs per month	Number	49.23
Net rent per month	Number	49.27

The feature values from this data source are then calculated by aggregating the data from all advertisments within a radius of 1,000 m around each household. In the case of numeric values, the arithmetic mean, in the case of logical values (e.g., existence of cellar or parking space), the relative frequency, and in the case of categorical variables, the relative frequency for each category is calculated.

Experimental data

Using the address information, geographic features from OSM were calculated using the offered Application Programming Interface (API) following the methodology described in Geographic features from OpenStreetMap data section. Features from real estate advertisments were computed as described in Feature extraction from real estate portals section. For the analysis in this paper, 34 features from the real estate portal data were present in more than 1000 advertisments and therefore used (the features include for example: category and type of the dwelling, rent per month, net rent per square meter, number of rooms, living area, year of construction, distances to next bus station / highway / kindergarden / school / shopping, existence of balcony, cable TV). Missing values have been encoded with − 1 being able to use the data in the further analysis.

Only for half of the customers (n=1,718), a sufficient number of five real estate advertisements was available (see Fig 3). This sample of customer data with a sufficient number of real estate advertisements \(\mathcal {B} \subset \mathcal {A}\) is used for the remaining analysis.

I assume that the sample of customers in \(\mathcal {B}\) is representative for utility customers and find evidence for that in the statistics presented in Table 3. The distribution of household properties are quite equal in both datasets, except the size and type of the residency: sample \(\mathcal {B}\) contains less larger dwellings and more apartments. By looking at the electricity consumption, there is a difference in the means of both datasets: M=12.24 kWh (SD=6.85) in \(\mathcal {B}\) and M=13.03 kWh (SD=7.02) in \(\mathcal {A}\). The difference mainly results from the fact that smaller homes and less houses are present in \(\mathcal {B}\). The electricity consumption of large homes with living ares above 145 m² show for example no significant difference (t(282)=−1.347,p>0.10). Therefore, one can assume that the sample is representative.

Classification results

The household classification methodology, depicted in Fig. 2, is applied using the Random Forest classifier with its implementation in the package ’randomForest’ (Liaw and Wiener 2015) for the statistical programming environment GNU R³. No automatic feature selection is done, since the Random Forest algorithm has a good internal ability to weight the importance of certain features. The aim of this study is to investigate, how well household characteristics can be recognized from different data sources. Following this explorative scope, explicitly neither a tuning of the Random Forest algorithm parameters nor a model selection (including test of other classifiers) is performed. Initially, a number of 500 trees is used (this is the standard value of this main parameter to control the algorithm in the used implementation) and a sensitivity analysis on how different values of this parameter affect the results is conducted.

The classification performance is evaluated using AUC as a well-known metric assessing the performance of classification models (Han et al. 2012). The metric quantifies the classification performance from 0 (lowest) to 1 (best), where 0.5 is considered as random classification. So, all values that are clearly higher than 0.5 or lower values are acceptable. For comparison with other studies and for better interpretation of the results, Accuracy and the harmonic mean of Precision and Recall, called F1-score (Hastie et al. 2009, chap. 7), is included in Table 3. To obtain a solid estimation of the classification performance, 10-fold cross-validation is applied. In this statistical technique, the data is randomly split into k=10 independent parts (so-called folds). Training and test is performed k times while in each iteration, one fold is used for test and the remaining folds for training. It is generally acknowledged that this technique gives a proper estimate of the classification performance, even with small datasets. In fact, the procedure even leads to a conservative estimate of the performance, given the fact that cross-validation slightly overestimates the variance of prediction results (Arlot and Celisse 2010). The classification performance results are presented in Fig. 4 together with 95%-confidence interval obtained using the Student’s t-Distribution (df=9).

One can see that all predictions made by the presented classification model are significantly better than random (AUC>0.5). Besides that, OSM data improves the classification performance significantly. This result replicates the findings from previous studies (Hopf et al. 2016; Hopf et al. 2017). The additional use of real estate portal data shows small improvements for the properties space and water heating, living area and the class > 5 persons. Especially the small class can be better recognized using the real estate portals.

On average, the use of both geographic data sources improved the classification performance in AUC by 12.8% (the improvements range between 8% for single households and 22% for dwellings > 145 m²), but especially the household type could be improved by 16.6% in both classes.

A very interesting result is that classification only based on OSM data can also lead to good results. The household type, for example, can be predicted better than using electricity consumption features. However, the prediction of household classes solely with OSM features leads to 14.19% lower AUC on average for all properties.

The results presented in Fig. 4 are obtained using 500 trees in the Random Forest algorithm. This is a typical parameter value and the standard in the used implementation. In an empirical study involving 29 datasets, Oshiro et al. (2012) document that larger numbers of trees do not significantly improve the classification performance, otherwise lower numbers of trees decreease the classification quality. To assess the stabilit of the models in this study and get an impression on possible improvements to the presented approach, a sensitivity analysis for this important parameter is conducted using n={10,100,250,500,1000} trees for the classification using all data available in this study. The results are shown in Fig. 5 as point estimates with a 95% confidence interval obtained using the t-Distribution from 10-fold cross-validation. For all binary properties (household and heating types) only one class is shown, because AUC results for the other classes are equal in each property. It is clear, that a low number of 10 trees leads to considerably lower classification performance. However, the model cannot be significantly improved by changing the number of trees between the values n={100,250,500,1000}. Thus, the standard value of 500 trees is quite a good choice in this presented case.

Discussion

The results from this study show that it is possible to considerably improve household classification models based on annual electricity consumption data by using freely available VGI data. Thus, the stated research question can be answered positively.

The results are in line with previous studies on factors that influence residential electric load (Kavousian et al. 2013; Heiple and Sailor 2008), which identify geographic data and household details as good explanatory variables for the energy consumption in households. So it is reasonable that geographic information – together with annual electricity consumption or alone – can serve as predictor for relevant details about residential energy customers, as shown in this paper. The real estate portal data has, despite the large amount of building related information contained, not helped much to increase the predictive quality.

Besides that, the presented results confirm also findings from a previous study with a similar household classification approach that was trained and validated with data from Switzerland and Germany, where also the transferrability of trained models (i.e., trainig of a model in one region and then applying this model to customers located in another country) was positively tested (Hopf et al. 2017).

Limitations

Findings in this paper are drawn from a dataset on utility customers from Switzerland. Thus, the concrete classification performance results may vary in other studies, but the performance gain from VGI should be replicable. In a recent study it could be, for instance, shown that household classification models can be trained with data from one country and applied to another one (Hopf et al. 2017). Due to the sparse number of real estate advertisements, not the complete number of households could be used in the study. The collection of further real estate advertisements, and also data from additional VGI data sources, would give a more comprehensive picture of the topic.

A possible limitation to the work is the aspect of privacy preservation and user acceptance of predictive information systems. Krishnamurti et al. (2012) and Mah et al. (2012) have studied the acceptance of end users for smart grid applications empirically and came to the conclusion that private customers have a positive attitude towards smart grid applications (consumption feedback, dynamic prices, etc.). However, it is necessary that utility companies educate their customers on the added value of smart grid technology and the transparency of data processing (Gangale et al. 2013) in order to place smart grid services well in the market. An important driver of customer acceptance are also financial incentives such as dynamic pricing of energy that are made possible by smart meters (Motsch 2012).

Future work

The phenomenon of VGI should be subject to future research. For instance, the persistence and reliability of VGI initiatives can be investigated considering knwoledge on the motivational factors and contributor behaviors in free/libre open source software projects (Stewart and Ammeter 2002; Crowston et al. 2007; Crowston et al. 2008), as a starting point. Besides that, other VGI data sources may be tested for their contribution in energy data analytics applications, especially data sources that contain event and environmental data (adding a time-dimension), such as social networks.

Considering the data quality of VGI data, the empirical definition of new features can help to overcome issues of data sparsity and variety. This is a critical issue and became visible in this paper with the large number of missing values in real estate advertisments. Possible methods to overcome the problem without discarding instances or feature vectors include the imputation of values (Saar-Tsechansky and Provost 2007), the inference of values from textual descriptions using text mining, or by treating all missing values in a vector as a separate class in model trainig and prediction. The application and comparison of such approaches is motivated to be a subject of future research.

Regarding the household classification artifact, the prediction models can be further enhanced by testing other parameter values for the Random Forest algorithm, benchmark additional algorithms with different parameter combinations, and automatic feature selection.

Finally, it would be interesting to investigate whether VGI data sources have also value for other topics in the field of Energy Informatics, such as load forecasting or studies regarding consumer behavior in energy markets.