Generating future weather files under climate change scenarios to support building energy simulation – A machine learning approach

Highlights

• A workflow to generate weather files under four climate change scenarios is proposed.

• Quantile-Quantile bias-correction is used to remove existing bias in GCM.

• A hybrid model of k-nearest-neighbour classification and Random Forest regression is used to downscale GCM data.

• Future weather files, year by year, under different climate change scenarios are constructed.

Abstract

General circulation models (GCM) have been used by researchers to assess the effect of climate change in different fields of study. In the case of building energy performance, GCMs can be used to evaluate future building energy performance through simulations. However, a key issue with the use of GCM data in building energy simulation is the inadequate resolution and bias of the data. Therefore, in order to use this data for simulation purposes and better predict future building performance, further processing is required. The first challenge is that the GCMs are usually biased, which means a considerable deviation can be found when the historical GCM data is compared to station observed weather data. The second challenge is that the GCM data has daily temporal resolution rather than the hourly resolution required in building energy simulation.

In order to utilize GCM data to estimate future building performance through simulation, the current study suggests a workflow that can be applied to climate change data. First, a bias-correction technique, known as the quantile-quantile method, is applied to remove the bias in the data in order to adapt GCMs to a specific location. The study then uses a hybrid classification-regression model to downscale the bias-corrected GCM data to generate future weather data at an hourly resolution for building energy simulation. In this case, the hybrid model is structured as a combined model, where a classification model serves as the main model together with an auxiliary regression model for cases when data is beyond the range of observed values. The proposed workflow uses observed weather data to determine similar weather patterns from historical data and use it to generate future weather data, contrary to previous studies, which use artificially generated data. However, in cases where the future GCM data showed temperatures ranging outside of the observed data, the study applied a trained regression model to generate hourly weather data.

The proposed workflow enables users to generate future weather files year by year under different climate change scenarios and, consequently, extreme weather characteristics are preserved for extreme or reliability analysis and design optimization.

Introduction

Historical data shows that Canada has experienced a warming rate of twice the global mean, while northern Canada has experienced triple the global mean [1]. According to the information presented by The Fifth Assessment Report of the Intergovernmental Panel on Climate Change (IPCC) Working Group I [2], over the period of 1880 and 2012, the global temperature has increased by about 0.85 °C. A Canadian study reported the annual temperature rise ranging from 0.5 °C to 4 °C for 16 major cities in Canada over the period of 1900 to 2013. For Montreal, the average annual temperature rose about 2 °C, where a 1.4 °C rise was noted during the summer, and a 2.7 °C rise was recorded in winter [1]. Ongoing temperature rising with respect to climate change and its possible catastrophic consequences on building performance have made more and more engineers and architects consider sustainability in building designs. However, as a first step, it is important to clearly understand the possible problem and consequences in order to offer appropriate solutions. In the building and construction industry, the topic of climate change and the possible impact on energy consumption of buildings have been studied by engineers and architects. Building energy simulation has been the main tool to evaluate the energy performance of buildings. In order to evaluate the building energy performance, meteorological data is required. Typical meteorological year weather files, which are used for simulations, reflect the historical weather condition and do not necessarily represent future weather data. In order to adequately represent future energy demand in building energy simulations, future weather data must be generated. The generated weather data must fit the purpose of the application. For example, in cases where the purpose of the building energy simulation is simply to estimate the future building energy consumption, a single future typical weather file representing a trend of multiple future years’ weather might be enough. However, one important aspect of climate change is the increase in the number of heatwaves, which lead to significant consequences, including heatwave related fatalities. As an example, during the summer of 2010 and 2018, heatwaves caused about 280 and 90 deaths, respectively, in the cold climate of Quebec, Canada [3]; of the 90 deaths in Quebec, 66 deaths occurred in Montreal, Quebec, Canada [4]. A survey conducted in 2011 showed that only half of the residential buildings were equipped with air-conditioning systems [5]. These facts might reflect the point that these buildings are not designed for extreme conditions, but rather, for typical cold climate weather conditions. This point directs the research toward creating future weather files that not only meet the intended initial use of finding typical building energy performance, but that also project the variability of performance for extreme conditions, including heatwaves. The generated weather files would enable architects, building engineers, and energy modelers to consider extreme future events at the building design stage in order to ensure that buildings and their associated energy systems could operate as expected under these conditions.

There have been different types of weather files, which are used to evaluate typical and extreme weather conditions using energy simulation, that are reviewed in a previous study [6]. There are many challenges in handling climate change data for building energy simulation, which are explained in the following section. Table 1 is nomenclature which explains the abbreviations used in this study.

General circulation models (GCM) have been used by researchers to assess the effect of climate change in different fields of study. GCMs mathematically simulate atmospheric, oceanic, and biotic interactions and combine them with radiative forcing scenarios to evaluate the future climates. The models consist of grid cells resulted from latitude and longitudinal divisions, in which the meteorological data is calculated [7]. Although these models help considerate the impact of climate change, the output data of the models can’t be directly used for building energy simulation. The challenges in applying the data are explained in the following section. Table 1..

GCM covers a vast geographical area, and the historical data based on the model is expected to deviate from the observed data of a specific location. Furthermore, the output of the circulation models has a coarse resolution, in terms of both spatial and temporal dimensions. The spatial resolution of the output from the circulation models is larger than 100 km × 100 km. As for the temporal resolution, typically, the output has a resolution of daily-average. This poses an issue for building energy simulation, which requires hourly meteorological data. In order to have a finer resolution, the data must be further processed. Finally, the GCM provides data for most weather parameters that are used in typical meteorological weather files, with a few exceptions, notably, dew point.

Moreover, GCM suffers from systematic bias, which means that there is a considerable deviation when the historical data of GCM is compared to observed data at stations. The GCM systematic bias, coarse resolution, and limited available weather parameters make it difficult to apply the output data, without further processing, for energy simulation purposes.

Therefore, in summary, there are three main challenges to use GCM data for building simulation:

• Coarse spatial resolution

• Coarse temporal resolution

• Bias in the model data

From simple to sophisticated methods, previous studies used various approaches to address these challenges.

Generally, there are two main approaches to process the GCM outputs; dynamical and/or statistical downscaling. Dynamical downscaling relies on further application of physics-based models for finer resolution outputs. Statistical downscaling, however, relies on the application of statistical rules and correlation for further processing. Each of these two approaches has pros and cons, which are discussed in the following section.

Depending on the availability of resources and expertise, different approaches are used to process the data in order to be used for building energy simulation. Some studies used dynamical downscaling, a method that requires an additional computationally intensive physics-based process based on a specific regional or local climate model. In this method, a regional or local climate model is nested in the GCM to use the GCM output as a boundary condition and take into account the hydrology, topography, and vegetation of the region to create finer resolution data. Kikumoto et al. [8] used GCM data as a boundary condition for regional climate models, namely, Model for Interdisciplinary Research on Climate (MIROC) and the Weather Research and Forecasting (WRF) to generate the future weather data of 2030s for a Japanese climate. The downscaled data was based on dynamical downscaling. For August, the sensible cooling load for a detached residential building increased by 15% from 2007 to 2030, mainly attributed to an average 1.52 °C rise in temperature comparing the two years [8].

Burger et al. [9] assessed the effect of climate change on the cooling and heating demand of an office building built during three different epochs: before World War I, after World War II, and from 2000 onward, in Vienna, Austria. They used REMO UBA regional climate model to dynamically downscale A1B climate change scenarios with a resolution of about 10 km × 10 km. They evaluated the results for two timeframes ranging from 2011 to 2040 and 2036–2065. The downscaled data was used for building energy simulation. In one case, the results showed an average 41% increase in annual cooling and 56% decrease in annual heating compared to the period of 1961–1990 as a result of a temperature rise of about 3 °C [9].

Although regional or local climate models provide finer temporal and spatial data, nesting of these models in the GCM models could be challenging as the time step and the grid resolution in the regional climate models are different from those in the GCMs. Moreover, physics-based calculations of finer resolution data would require expensive computational resources. In addition, when using dynamical downscaling, the bias of GCM data may not be removed completely, and further processing may still be required. Due to the mentioned restrictions, statistical downscaling techniques have been used to process the GCM data for various applications in different fields.

Other studies used statistical downscaling methods. These methods rely on the availability of observed weather data and can be used for all scenarios of climate change. Statistical methods provide point-scale climatic parameters and can be applied to regional and/or global models. In the statistical approach, a critical assumption made is stationary, which means that, while the climate changes, the statistical relations among the meteorological parameters remain constant over time [10], [11].

Generally, statistical downscaling is categorized into three main groups: linear methods, weather generators, and weather classification. The first two groups have been used for building simulation in previous studies and are briefly explained in the following sections, while the last group – weather classification, is one of the subjects of interest in this study.

Linear methods, including the delta method (also known as morphing in building simulation) [12], are easy to use and are widely applied in previous studies. In morphing, a changing factor is calculated by comparing the daily values of historical and future data and then applying it directly to the hourly-observed data to achieve future hourly data. These factors can be additive or multiplicative or a combination of the two depending on the weather parameter. For example, the downscaling of atmospheric pressure is additive (adding the change of means to each value for all hours), and the downscaling of the wind is multiplicative (multiplying the mean change to each value for all hours). The procedure is done separately for each weather parameter and, therefore, the correlations among the weather parameters are ignored. Therefore, the morphing method can change the mean and variance of the historical data to the GCM data, but other statistics of the data such as the 25th or 75th percentile of the data do not necessarily change. In other words, the morphing method might transfer the intensity of the GCM data to the downscaled data, but it may not transfer the frequency.

Due to its simplicity, the method is widely used to predict the future climate condition, including in American climate [13], [14], [15], [16], Canadian climate [17], [18], [19], Swiss climate [20], Swedish climate [21], [22], Spanish climate [23], Italian climate [24], British climate [25], and Chinese climate [26].

The morphing method is designed to use a TMY file, and the same is applied in these studies as well. The TMY is generated from historical data with extreme events removed [11]. The application of morphing on a single TMY might lead to project a single annual building cooling or heating energy consumptions for a long period of time. The mentioned feature makes the morphing method suitable as long as the goal of the application is to estimate the building energy consumptions over a long period of time. However, in cases where multiple weather years are required, the morphing method may not be suitable. There are some tools that use the morphing method to downscale the future GCM weather data, such as “CCWorldWeatherGen” or “WeatherShift™ tool”, which are used in building simulation [27], [28], [29]. One important point in using these tools is respecting the time frame of the base TMY file as it should be between the years 1961–1990 for the former and 1976–2005 for the later tool; otherwise, the output would be overestimated.

When the application is to optimize the design e.g. robust design, reliable design, or other design aspects, multiple weather years, are required. This is because of the fact that these design methods rely on statistics requiring multiple samples. For example, for designing a weather-robust building, providing multiple weather years would be an essential requirement for design optimization. As another example, for reliability-based design, extreme events must be taken into account. However, these events are disregarded by the morphing method on TMY, as it ignores the frequencies in the processing of data (e.g. frequencies of multiple warm days/heat waves) [20]. Moreover, using the morphing method may not remove all the existing biases between the GCM historical data and the observation data. As such, weather generators are also used to generate multiple future weather years stochastically to capture the yearly variations over a long period of time.

Other studies used weather generators to evaluate the effect of climate change on buildings [26], [30], [31], [32]. Weather generators are used for temporal downscaling. These statistical models generate numerous possible time-series weather parameters using statistics of several-years historical data of weather parameters applied to the GCM model output. Due to the stochastic nature of data generation, the generated data might change when the process is repeated and may require further processing to make a single weather file. The data generated from the generators are different from the observed data and only keep the statistical characteristics of the mean and variance of the GCM data. Therefore, all the generated data is artificial data that synoptically have the same statistics as the GCM data. Among the weather generators, only a few are able to consider the relationship between the weather elements when multiple parameters are predicted [11], and they are able to generate only a few weather parameters such as temperature and solar radiation [31]. Furthermore, thousands of different generated data is produced by weather generators. Therefore, in order to adapt the data for building energy simulation, further processing is required [26].

To prepare the generated data for extreme analysis, Nik [22] suggested creating a typical downscaled year (TDY) together with one extreme cold year (ECY) and one extreme warm year (EWY). In order to create these weather years, the same procedure of making a typical meteorological year (TMY) is used, except that only the temperature variable is considered in the procedure e.g., one year with all months having typical temperature values (TDY), one year with all months having low-temperature values (ECY), and one year with all months having high-temperature values (EWY). The reason behind considering only the temperature was stated to be “the difficulties and uncertainties in weighting the climatic variables”. The same method was used in a newer study to downscale regional climate model data in order to consider typical and extreme conditions for the projection of weather data for different future timeframes, namely, 2010–2039, 2040–2069, and 2070–2099 [20]. Although the method reduced the required number of simulations for building design optimization, the method seems to create artificial extreme weather years that may never happen in the future, especially without the bias-correction of the RCM data that might overestimate the extreme conditions.

The issue of the weighting of the climatic variables mentioned by the previous two studies [20], [22] is recently addressed with a novel method in the previous study conducted by the authors for building energy simulation application [33]. The method will be briefly explained in this study as well.

From the literature review, it is perceived that the morphing method may not be used for design optimization of the building and HVAC systems as it ignores the frequency of data, and therefore, it ignores the extreme conditions such as heatwaves in summers. Moreover, the morphing techniques are applied to each of the weather parameters separately and consequently ignoring the correlations among the weather parameters.

Weather generators, on the other hand, may be used for design optimization; however, the method generates a large amount of artificial data that only maintains the mean and variance of the GCM data for each weather parameter separately, ignoring once again the correlation between the parameters. The literature review showed that previous studies mostly used morphing or weather generators. In other industries, weather classification or weather typing schemes is another method used for downscaling the GCM climate change data. However, due to certain limitations, this method has not been deployed in the building industry.

Weather classification relates a class of future weather patterns to locally observed weather data, and the future weather data are synoptically selected from the observed weather data. In this method, the effect of climate change is estimated by evaluating the frequency and intensity of the change of the weather pattern parameters from the GCM output. In other words, the future GCM data is compared to historically observed data, and a class of historical data is selected according to statistical resemblance. The selected historical weather pattern values will then represent the future climatic weather condition. The method can be used for normal and non-normal weather parameters such as temperature and wind speed. However, a large set of historical observations is required [11]. This method is more sophisticated than the simple morphing method in the case of data analysis but can create real data selected from the past and therefore keeps the real relationship among the weather parameters in hourly values.

In contrast to the morphing method, this method can transfer all the information of GCM data (including percentiles) to the downscaled data. This method has been used in hydrological studies [34], [35], but there is a lack of literature in applying the method to building energy performance. In the building energy simulation industry, weather typing or classification is used to create a representation of the long-term (usually 10–30 years) weather pattern from the historical data using a typical meteorological year. The Sandia method, a method for generating typical meteorological year files, can be categorized in this group. However, the method is not used to downscale the future GCM data. The main disadvantage of this method is its inability to predict new values that are beyond the range of the historically observed data (increase in intensity). Building energy consumption is highly correlated with outdoor air temperature, which is expected to increase in the future unprecedentedly. Therefore, the weather classification alone may not be used to downscale the future climate data for building energy consumption. In this study, a hybrid classification-regression model is proposed to downscale climate change GCM data.

Weather classification can retain all the statistics of the GCM data; however, it is incapable of keeping the intensity of data when a new out-of-range of historical data is in the GCM data. Therefore, the combination of the weather classification with a regression model as a hybrid model can be a promising method to downscale the GCM or RCM data to be used for building energy simulation. The method can partially apply an algorithm that is currently already used in the building simulation industry and uses real historical weather data rather than artificially generated values. The regression model will be trained using observed historical data and will be used only for conditions when the GCM data is higher or less than any previously experienced data. The hybrid application of this method can transfer all the statistics of the GCM data while keeping the intensity of the data as well.

There is a shortage of studies (if any) that apply weather classification as a method to downscale the GCM data to be used for building simulation. The reason for that, as mentioned before, is that the building energy performance is highly dependent on outdoor air temperature, which is expected to increase in the future; the classification model is not properly downscaling the new values that are higher or lesser than any observed values. Therefore, a regression model will be combined with the classification model for new unseen values larger or smaller than the observed values. This method is introduced in this study as part of the workflow.