Now that we are aware of the potentially devastating consequences of global warming, one of the top priorities of modern societies around the world should be to reduce the emission of greenhouse gases. Since most of the electricity we consume is still produced from ‘dirty’ sources, such as coal and fossil fuels, one way to reduce our carbon footprint is to keep an eye on the power we use at home—our residential energy consumption (REC). 

Scientific studies have shown that the largest contributors to REC are heating and cooling in most places in the world. Thus, one could safely infer that REC is strongly related to a region’s temperature. Such assumptions have been the basis of many mathematical models that seek to explain the relationship between temperature and REC. However, the truth seems to be much more complex than this, and there are many additional factors that have to be considered to produce better models.

 

“Previous studies warned us that global warming will increase REC during summer because higher temperature drives up cooling demand. However, many climate-related power outages in recent years happened in winter times during strong snowfall events. This got us curious about the role of precipitation in shaping REC,” explains Cuihui Xia, lead author of the study from Institute of Tibetan Plateau Research, Chinese Academy of Sciences. In the latest paper, Xia and colleagues took an innovative approach to understand how different climate-related factors can affect REC in Tibet, a cold alpine region with distinct rural and urban areas. The researchers decided to leverage machine learning and available REC and climate data from Tibet to train multiple models and compare their performance. 

 

In general, machine learning models, once trained with real-world data, act as ‘black boxes.’ They have good predictive accuracy without making any assumptions beforehand, but there is not much one can do to understand how each of the input factors (such as temperature, precipitation, income and population) is ultimately weighed by the system to predict the outcome (REC). This low interpretability is a huge drawback because science aims to not only make correct predictions, but also explain the underlying reasoning. To tackle this limitation, Xia and colleagues employed the latest techniques put forth by the interpretable machine learning community to peek inside these ‘black boxes.’

 

This approach led to some interesting findings. The results show that precipitation during the cold season is a major contributor to Tibet’s REC. This was especially so in days where temperature was below freezing, which produces snowfall instead of rain. Thus, it is important that scientists attempting to understand the impact of climate on energy consumption (and greenhouse gas emissions) don’t chalk everything up to the variable of temperature. “Our study highlights that climate-related energy studies should treat climate as a whole rather than focusing on temperature alone, particularly for studies concerning cold winters and extreme weathers such as blizzards,”added Xia. 

 

Another important finding was that there were marked disparities in the way REC was shaped by climate between urban and rural areas. Urban areas, in cold and snowy days, were more susceptible to increasing their REC compared to less-wealthy rural areas. The capacity to make this sort of comparisons between urban and rural areas while considering socioeconomic factors is another advantage of the data-driven approach Xia’s team used to produce their interpretable models.

 

Editor's note: This study was supported by STEP, a TPE-related science project