The review paper of Swan and Ugursal (2009) outlines a clear classification of available models that estimate heat demand of buildings. Bottom-up models have the advantage that they estimate the energy demand of small consumer aggregates, as for example buildings that can be linked to geographical coordinates. To estimate the energy demand at the building level, the two main alternatives are to use either engineering or statistical models.

Engineering models are based on building physics and require detailed information on the building geometry as well as the envelope components and their associated U and g values. At the national scale such detailed data are not available and default values for many parameters must be used (Perez et al., 2011) when simulating the heat demand of a large building stock. Choosing this approach implies that buildings of same type have the same losses by conduction through the building envelope. Energy demand, therefore, only depends on local climate conditions and solar gains.

Statistical models encapsulate the complex set of building characteristics that explain the energy demand in a smaller set of parameters. Regression models estimate these parameters using calibration samples of buildings with known energy consumption. Another advantage of basing the model on measured energy demand is that it includes the influence of the users (inhabitants and facility manager). This influence often leads to a difference between calculated and actual consumption even in cases where the characteristics of the building are known in detail (De Wilde, 2014).

These reasons motivate the choice of a statistical extrapolation method based on measured final energy consumption; such a method accounts for the behavior of the user and the variability of demand between buildings with similar characteristics. A representative sample of around 27,000 buildings containing measured final energy for SH and DHW is used to calibrate the model.

There are numerous methods to simulate load curves. The review paper of Grandjean et al. (2012) classifies the most relevant algorithms into five groups, including both random (replicating the diversity of consecutive days) and deterministic models (replicating average behavior). For example, high-resolution models, such as Marszal-Pomianowska et al. (2016), seek to include the stochastic variability of electricity demand related to the behavior of the inhabitants. At the Swiss national level, the goal is rather to provide a large-scale model based on the average behavior of users and their related activities. The proposed approach is closer to that of the FORECAST/eLOAD model (Jakob et al., 2014). The primary target is to offer the Swiss municipalities a tool allowing to identify electricity saving potential and to predesign an ECO21 type program. The key question is how the electric demand of a municipality is divided into the various activities and electric appliances.

The model presented in Section “Geo-Dependent Electricity Demand” is the result of a collaboration between the academic research project SCCER FEEB&D (2017) and the local utility SIG. The basis for this model is a study of Le Strat (2011) that characterizes the electric consumption of a territory including all the energy service companies (ESCO) of the EOS holding group. This model decomposes the electricity demand into 36 activities and 18 groups of electric appliances. The planned outcome of this collaboration is the ElectroWhat web platform, which seeks to answer the question “who consumes where, when and for what use” at the national level. Throughout this paper, the term ElectroWhat often also refers to the geo-dependent electricity demand model.

Finally, the resulting GIS data of both models is stored in a database developed in collaboration with EPFL in Lausanne, also containing the renewable PV rooftop potential (Assouline et al., 2015) and low enthalpy geothermal potential (in preparation). Such a database offers the opportunity to perform many territorial studies at the Swiss national level.

The chosen approach, presented in this section, is a statistical regression bottom-up model (Schneider et al., 2016). It is based on measured heat demand and heated surfaces values of a representative set of around 27,000 buildings. It estimates the heat demand per square meter of each building of the Swiss national building register as a function of its category, age, and location, and its heated surface as a function of the category, gross surface, and/or dwelling surface. A bootstrap resampling algorithm quantifies the uncertainty inherent to the use of average demand values per age and category. This model improves previous work done by Eicher et al. (2011). Figure 1 outlines the general structure of the model. The heated surface A_E (square meter) and the floor specific heat demand E (MJ/m²/year) are estimated combining basic information provided by the Swiss building register (OFS, 2014a) with average statistics calculated using calibration sets. The estimations of the heated surface A_E and floor specific heat demand E are done separately, since they allow to take account of regional climate conditions that may vary a lot depending on geographical position. We assume that only E depends on climatic conditions, whereas the relation between A_E and gross building surface (A_gross) and/or dwelling surface (A_dwelling) do not. The final energy demand SE_B (MJ/year) of a building is obtained by multiplying these two estimations.

Finally, for each pixel of territory, the estimated heat demand is summed over all buildings and a bootstrap algorithm allows to compute a confidence interval around the estimated value given by the model.

This section describes the model used to decompose the electric demand of a territorial unit into various activities and groups of electric appliances (Schneider et al., 2017). The choice of a Swiss municipality is adequate since it represents the smallest political entity of Switzerland. For this reason many socioeconomic indicators are available at this level. In 2016, Switzerland had 2,287 municipalities of quite heterogeneous size. The number of residential dwellings ranges from 14 to 224,774, the city of Zürich being the biggest municipality after the city of Geneva. Since the spatial constraint for electricity is lower than for heat, it did not seem of interest to break down the territory into pixels as in the case of heat demand. Another particularity of certain regions is that the ratio between main and secondary residences varies a lot. The percentage of main residential dwellings ranges from 52% for the Canton of Graubünden to 98.8% for the Canton of Basel. This socio-economic characteristic has an important influence on energy demand.

As illustrated in Figure 5, the yearly demand of a municipality is split into three main sectors, each one with its own estimation algorithm as described in the following two subsections. The yearly consumption (YC) is split into 36 activities and 18 electric appliances such as lighting, fridges, TV, etc. A further step consists in transforming the yearly demand into estimated load curves using a library of load curve shapes.

The heat demand model of Section “Geo-Dependent Heat Demand” allows to estimate the final energy demand for SH and DHW. Combining these estimations with fractions of dwelling using electric heat production systems permits to estimate the amount of electricity used for SH and DHW. In case electric heat production is based on a heat pump, we use an average coefficient of performance of 2.5 to convert useful heat into electric demand.

Relational database servers and data exchange protocols, as for example the simple object access protocol (SOAP) are powerful tools allowing data interconnection originating from different sources. To facilitate data exchange, a database has been set-up in collaboration with EPFL that consolidates results of the geo-dependent energy supply and demand models. This consolidation fulfills the following requirements:

Gives the results of our model estimations for energy supply and demand on a harmonized geographical aggregation level (municipality level and 200 m × 200 m pixel size).

The actual release of the database contains rooftop PV potential (Assouline et al., 2017) with monthly and yearly values for each municipality in Switzerland as well as yearly heat demand and electricity estimations derived from the models presented in this paper. The database provides PV supply potential and heat demand estimations up to a resolution of 200 m × 200 m pixels. Concerning electricity demand, the database contains for each municipality the yearly demand estimation split up into activities and appliances. A dedicated web-service allows to query the database using the SOAP protocol. The user manual and sample of scripts to query the database are in the report of Schneider et al. (2015).

Table 2 contains aggregated values—per building category and construction period—of the total heated surface A_E and final energy consumption for SH and DHW. Sufficient information is available to calculate these estimations for around 90% of the 1.93 million buildings listed in the 2013 version of the GWR. These values lead to a total heated surface estimation of 702 million m² and a total final energy demand of 93.7 TWh/year (76.8 TWh/year for residential and 16.9 TWh/year for mixed use and non-residential buildings). Within the residential sector 71% of the energy is consumed by buildings built before 1980. The fact that these buildings also have a high heat demand per square meter emphasizes the importance of thermal envelope retrofitting.

Comparing these numbers with the estimation of an alternative study permits to detect a possible bias of the present model. For year 2013, Kemmler et al. (2014) estimate SE_CH to be 91.1 TWh/year, (59.5 TWh/year for residential and 31.6 TWh/year for mixed use and non-residential buildings). Although the two models have a good concordance for global demand, significant differences exist when considering the residential and non-residential sectors separately, mainly due to the differences between the estimated heated surfaces. In the case of the residential buildings, our model estimates a heated surface of 580 million square meter that is significantly higher than the one 509 million square meter estimated by Kemmler et al. (2014). This lower estimation would lead to a ratio f _dwell of 1.2 that is significantly lower than the value of Table 1.

Since the model delivers energy demand estimation at the building level, Eq. 9 permits to compute the aggregated demand for pixels of any size. Dividing the Swiss territory into a raster of pixels and exporting, pixel by pixel, the demand to an information system such as QGIS, allows us to draw heat demand maps. Figure 6 gives an example of such a heat density map for Switzerland, with a zoom on the town of Martigny. A color ramp indicates the final energy consumption per square meter of territory for each pixel of 200 m × 200 m size. Such maps permit to identify portions of territory having high demand density, being consequently suitable for district heating deployment. Cross-matching these demand maps with other geo-referred information, as for example availability of waste heat, permits to estimate the exploitation potential of such resources. Other information, as for instance, the percentage of demand covered by fossil energy carriers can be added to derive estimated CO₂ emission maps.

The previously presented geo-dependent heat demand database allows estimating the savings that could be achieved by large-scale retrofit of building envelope. The new energy policy scenario (NEP) of the Swiss Energy Strategy 2050 targets reducing today final energy demand for SH and DHW production to around 42 TWh/year; this represents a decrease of 53%. This section challenges this scenario by making the optimistic assumption that today retrofit rate will double to reach 2%. If current retrofit practices are applied, the SIA 380 standard (SIA, 2009) sets an upper limit Q_SH,lim (MJ/m²) for the calculated useful energy demand per square meter. This demand is calculated under normed conditions for indoor temperature and ventilation rate that are not often fulfilled in real use conditions after retrofit. This discrepancy between the normed and real use of the building induces a performance gap. This gap is the difference between the theoretical savings QS_Gross (MJ/m²) defined by (12)QSGross=QSH−QSH,lim, and the actual savings QS_Net (MJ/m²) defined by (13)QSNet=QSH−QSH_AR, where Q_SH,AR is the actual measured energy demand per square meter after retrofit. The analysis of ten representative case studies (Khoury et al., 2016) of recently retrofitted post-war multi-family residential buildings allowed to characterize the performance gap by means of a statistical correlation between QS_Gross and QS_Net (14)QSNet=0.0009(QSGross)2+0.17QSGrossifQSGross<600MJ/m2.

This relation (R² = 0.9897) takes account of the entire retrofit process, from the design stage (choice of solutions and use of simulation software) to the use of the buildings by the occupants and energy managers. It can be viewed as a characterization of the current retrofit and operation practices. In case QS_Gross exceeds 600, we use the more conservative value QS_Net = 0.7 QS_Gross. This rule applies to buildings having a theoretical energy savings that lies outside the range of the building sample analyzed by Khoury et al. (2016).

The analysis of this section will show that taking account of the performance gap has a significant effect on the savings of large-scale retrofit programs. Figure 7 shows the procedure used to estimate the net saving potential of a group of buildings that are supposed to be retrofitted in accordance with the efficiency limit of the SIA 380 standard. We start by assigning to each of these buildings a shape factor, which is defined as the ratio between the building thermal envelope surface A_th and its heated surface A_E. These ratios are taken from the SIA 2031 standard (SIA, 2016) for all the building categories, except for the multi-family residential buildings. For the latter, the factors are estimated on the basis of the number of floors using the relation below, which was derived from the GEAK database.

These shape factors permit to compute for all buildings the limit value Q_SH,lim, which is the heat per square meter demand a building should not exceed after retrofit. The heat demand of the building for SH before retrofit is estimated by a randomly picked Q_SH value taken from our calibration set of buildings of same category and construction period (see Eq. 5). This randomization replicates the variability of the observed distribution of heat demand of buildings of same category. Using an average value instead of this distribution would not have any effect on the average QS_Gross value, but would bias the QS_Net estimation, because of the quadratic term of Eq. 14. Putting these two values into Eq. 12 allows to estimate the gross saving potential per square meter QS_Gross for each building. Eq. 14 allows to compute the net saving potential per square meter that is then converted into final energy using Eq. 16.

Finally, to compute the total net savings ESNetCH, we sum the net saving potential ES_Net of all buildings supposed to be retrofitted. This procedure is repeated 1,000 times and the average of the ESNetCH values gives the net saving potential of the retrofitted building group.

For future projections concerning the building stock size and the evolution of its energy demand, we refer to the report of Prognos (Kirchner et al., 2012). This study expects an increase of heated surface of 32% between 2010 and 2050. In the present simulation, it was supposed that all new buildings added to the current building stock consume 100 MJ/m²/y of final energy for SH and DHW.

Figure 8 shows how the possible evolution of the final energy demand for 2020, 2035, and 2050 using the previously explained estimation procedure. Under the assumption no retrofit is done, the total demand SE_CH will increase proportionally less than the additional heated surface, since the new buildings are supposed to be very efficient. The doted gray lines are the expected evolution of demand according to the new energy policy scenario (NEP) and business as usual scenario (WWB) of the 2050 Swiss energy strategy. In our 2% per year retrofit rate scenario, it is assumed that the buildings offering the highest saving potential are retrofitted first. Even following this optimistic assumption the expected savings ESNetCH are significantly lower than the projection given by the WWB scenario. Disregarding the performance gap leads to savings ESGrossCH that are close to the projections of Prognos.

This section presents two kinds of validations done by comparing estimates obtained from the model with available electricity consumption statistics. A first validation consists in comparing estimated and actual annual consumptions at the national and at different regional levels. This enables to check the accuracy of the unitary consumptions used to estimate the YCs in the various sectors of activity. A second validation compares the estimated hourly demand with the measured one. The way the model builds the estimated hourly load curve implies that load curves are identical for all the days of the same month and day type (working, non-working). It is thus adequate to compare the estimated load curve with the average load computed by taking separately working and non-working days for each month and hour of the day. This kind of comparison checks if the shapes of the load curve library and the weights used for aggregation replicate actual demand.

Figure 12 shows an example of a GIS map combining yearly demand per capita, distribution of demand into four main activity sectors and total demand for some municipalities near the city of Lausanne. A quite high heterogeneity exists among the shares consumed per sector, also influencing the per capita consumption. In the municipalities of Aclens and Mex a large part of the demand is attributable to the services and industry sector, which induces a high demand per capita compared to a residential municipality such as Bussigny. This difference of socio-economic structure influences the distribution of demand into electric appliances and the shape of the load curve.

The Electowhat model is used to set-up a prototype web platform allowing to explore the consumption specificities of each Swiss municipality. A report containing the main indicators is enriched with charts illustrating the distribution of the yearly demand into activities and electric appliances. This detailed distribution of electric consumption can be downloaded in Excel format for further treatment. A future step will be to include on this platform the experience gained by SIG during the implementation and monitoring of their electricity efficiency program ECO21. The platform will use the decomposed electricity demand provided by the model to help the user to design an efficiency program of ECO21 type in any region of Switzerland.

The geo-dependent heat demand model presented in Section “Geo-Dependent Heat Demand” is an alternative to commonly used calculation models based on building physics, which suffer from a lack of accuracy when used at large territorial scale, due to the limited available information on the building characteristics. Calibrating the model using measured final energy consumption bears the advantage of replicating the difference in energy demand between buildings with very similar characteristics. An innovative approach consists in using a bootstrap algorithm to compute a confidence interval in addition to the estimated demand of a pixel of territory. These confidence intervals quantify the part of variability the model cannot reproduce, when the specific heat demand Q_h of a building is replaced by the average demand of the buildings of same category and age.

There are two important assumptions linked to this kind of approach. The first is that the GWR database contains the majority of the heated buildings of Switzerland. The second is that the samples used to calibrate the bottom-up model are representative. Important deviations from these assumptions would introduce bias in the estimations. One possibility to detect bias is to compare, for example, the aggregate demand of all buildings with other available national scale statistics. As shown in Section “Heat Demand Statistics Aggregated at the National Level,” the total aggregated final energy demand for heating is 93.7 TWh/year (76.8 TWh/year for residential and 16.9 TWh/year for mixed use and non-residential buildings). Total demand is in good agreement with an alternative estimation based on national statistics, but significant differences exist when considering the demand of residential and non-residential sectors separately. The main cause for this discrepancy is the differences between the estimations of heated surfaces. Since the f _dwell values of Table 1 are based on a large number of observations, we are quite confident with our larger heated surface estimations for the residential sector. The alternative estimation used for comparison also comes from a model, which is subject to its own uncertainties.

The actual consumption data for buildings contained in the SITG IDC database allows to identify 206 pixels of 200 m × 200 m with known total final energy demand. Estimated demand together with its confidence interval and actual demand are compared for each of these pixels to test the accuracy of the bootstrap algorithm. For three chosen confidence levels α = 10, 5, and 1%, the proportion of pixels having their actual energy demand value outside the confidence interval are 15.5, 8.3, and 2.4%, respectively. The fact that these proportions are close to α, show the adequacy of the bootstrap algorithm.

Since retrofit has a major role to play in reducing demand and achieving the goals of the 2050 energy strategy of the Swiss Federal council, it is interesting to challenge the projections made by Prognos (Kirchner et al., 2012). Section “Impact of Large-scale Retrofit on Heat Demand” gives an example of how the heat demand model can be used to estimate the achievable heat demand saving potential of the Swiss building stock. For this sake, we relied on previous results (Khoury et al., 2016) concerning the performance gap between the gross and net heat demand savings achieved by building retrofit, which reflects the current retrofit and operation practices. When applying the results of the above mentioned study to our model, we estimate that if the entire Swiss building stock undergoes deep energy retrofit, the gross saving potential for SH would amount to 38.1 TWh/year. When taking into account the performance gap, the net saving potential for SH reduces to 18.4 TWh/year, i.e., almost half of the gross saving potential. A similar ratio between net and gross saving potential was already observed at a smaller scale in the canton of Geneva. Comparing projected savings for years 2020, 2035, and 2050 with the goals of the energy strategy 2050, we see that even when making optimistic assumptions on retrofit rate and selection of buildings, the net estimated savings remain far from the anticipated targets. This result is based on the assumption that the relation between net and gross saving potential applies to all building categories. Even if the projection of Figure 8 deserves deeper investigation, this first approximation already raises interesting questions. The projections of Prognos do not take into account the performance gap and may for this reason be too optimistic.

The approach given here has some limitations that are due to data availability. For instance, having an exhaustive GWR register would answer questions concerning possibly missing heated surface. A larger calibration set uniformly covering the entire Swiss territory and all building categories would also increase the accuracy of the model, as well as reducing the size of the confidence intervals.

The accuracy of the geo-dependent electricity demand model strongly depends on the detail level of the data available to calibrate the unitary consumptions per activity. This is illustrated in Table 3 that compares the estimated versus actual total aggregated demand for Cantons where local demand data was available—or not—for calibration. In the case of the Canton of Geneva, detailed consumption per activity is known, leading to representative unitary consumption of this region. In the Canton of Zurich, average Swiss values are used based on national energy consumption surveys for various activities. Although these default values lead to an unbiased estimation at the national level, they cannot replicate in detail the variability that energy demand may have between different regions, within the same activity. This phenomenon is also observed on a smaller scale, as shown in Figure 9, showing the errors at the level of each municipality, even when the model is calibrated with accurate Cantonal unitary consumptions. This reflects the fact that electricity demand within a same group of activities has a quite high variance that a large-scale model cannot reproduce.

A second type of validation compares estimated and actual load curves to check if the aggregation of the individual estimated load curves per activity and electric appliance lead to an overall demand, which is close to the average demand of a representative working and non-working day. Again, the added value of having local calibration data available is highlighted by the fact that the model replicates more accurately the demand of the Canton of Geneva than the overall national demand. Unlike the algorithm used for the FORECAST/eLOAD platform, no adaptations are done after the comparison. In our case the validation only serves to check if the load curve library and the weights used to combine the individual load curves reflect the average actual demand. Making further adaptations based on least square difference minimization may lead to completely biased coefficients since the number of degrees of freedom of the model is larger than the number of measured observations. We choose to let the model rely on parameters based on case study measurements and time of use model estimations for household appliances, without making any a posteriori adjustments.

It is interesting to observe that the fraction of heat production covered by electricity varies considerably according to the considered region (from 1% for Basel up to 25% for Vallis). This has an effect on the seasonal variation of the electric load curve of the region under consideration. The canton of Geneva has, for example, only a small fraction of electric heating (around 5% of total final energy). Consequently, the seasonal effect is less marked on Figure 10 than in Figure 11, since other regions of Switzerland have higher shares of electric heat production systems.

The proposed bottom-up model operates at the building level and allows heat demand estimations for every geographical aggregation level. It is calibrated using large sets of measured final energy demand. This data was used to compute heat demand statistics per building category and construction period. Basing the model on measured consumptions has the advantage of including the influence of the user. In addition, this approach allows to estimate confidence intervals by means of a bootstrap resampling algorithm. The aggregated overall demand for the entire Swiss territory is in good agreement with national statistics, speaking in favor of the model.

A first outcome of the model is to build a database providing the heat demand for a pixel grid of the entire Swiss territory. As second possible use of the model, is the estimation of expected savings achievable through retrofit (see “Impact of Large-scale Retrofit on Heat Demand”). It is interesting to notice that taking account of the performance gap significantly reduces the projected savings (of approximately 50%). This kind of observation was already pointed out by Khoury (2014) in his study on multi-family residential buildings of the Canton of Geneva. Such discrepancies reveal key questions on some of the energy savings projections made for the 2050 energy strategy of the Swiss Federal council.

The used model breaks down the total estimated electricity demand of a municipality into estimated hourly load curves per activity and electric appliance. The University of Geneva, in collaboration with the local utility SIG, is currently updating the model in view of simulating the year 2015. The model—today calibrated using the year 2008—replicates well the average behavior at large-scale, but of course cannot replicate all local specificities. This is a consequence of the fact that the chosen groups of activities have some heterogeneity concerning the consumption per employee. For example, consider the activity “manufacture of machines and tools,” for which the energy intensity varies a lot depending on the type of manufactured good. Some regions might be more specialized in one or the other type of good, having consequently differences of consumption per employee. The consequence of using Swiss average values for large portions of the Swiss territory is that a part of the variability of demand remains unexplained, as seen in Table 3 and in Figure 9.

The comparison of aggregated estimated load curves with actual ones shows that the model replicates the most important seasonal and intraday variations. The present simulation can be done for any geographical level for which the required input variables reflecting socioeconomic activity and building stock are available. By default, the municipality level should be used, since it is the smallest geographical unit having all the required input information easily available at the national scale. This model also opens the perspective of making ex ante estimations of savings that could be achieved by standard efficiency measure packages. To this end, in collaboration with the SIG, it is planned to set-up a web-platform helping the user to design an efficiency program similar to ECO21 in any region of Switzerland.

Finally, the results of the geo-dependent heat and electricity models are consolidated with geo-dependent renewable resource data. This database is shared with other research partners and serves as base input to assess energy transition scenarios. Between 2018 and 2020, the second phase of the SCCER FEEB&D project (SCCER FEEB&D, 2017) will explore how to find the best trade-off between energy efficiency (building retrofit, ECO21 type programs, etc.) and use of renewable resources (PV, geothermal, heat pumps, waste heat, free-cooling, etc.). Such models combined with a consolidated database are the foundation for exploring energy transition pathways and deriving policy recommendations.