This study has focused on many aspects related to heating systems, heat demand forecasting and data mining using SSAS. However, there remain many more unexplored aspects some of which were already discussed in the previous chapter.
This final chapter discusses and summarizes what future studies should focus on and also discusses the future of heat demand forecasting in association with emerging trends.
6.1 Recommendations for future studies
A more thorough investigation related to the mining models and changing the parameters of these models should be conducted. Tweaking the parameter values can well improve the accuracy of the forecasting and more inputs are possible.
Some buildings have scheduled heating programs for a day-time and a night-time.
If these patterns are known and set as an input to the models, the accuracy can improve significantly. For knowing that the indoor climate conditions are optimal, more sensors would be required in the buildings. This way we would have some understanding if the building is over or under heated. Knowing that the building is really hot, there is less need for heat demand and vice versa. This would also act as a good input for the mining models.
Other things to consider in the future works are included in the following list:
1. To generalize the results, there should be more case buildings to compare the results of the data mining models.
2. In order to understand the difference between the forecasted and the observed weather values, a comparison of forecast model (using forecasted weather values and observed weather values as inputs) accuracies should be conducted.
3. Summer season could be left out and run the mining models only during the winter, spring and autumn seasons. During the summer season in
94
Finland, the heat demand is considerably smaller than during the other seasons.
4. Hot tap water demand could be separated. Now the hot tap water and the space heating are both included in the total heat consumption data. The use of hot water depends more on the occupant behavior in residential buildings.
5. The performance of the models was not measured systematically although models with fewer inputs were evaluated for accuracy. These models with fewer inputs are likely more cost-effective.
6. When there is enough building data, e.g. from hundreds of buildings, it is possible to use this as the base for the forecasting. Similar buildings in similar weather conditions are likely to have similar requirements for heat demand.
7. The built deployment model for BEMS should be evaluated in real action and expanded as necessary.
6.2 About the emerging trends
When discussing the future, it is also important to understand bigger trends and cross-disciplinary development. These trends include the already discussed Big Data but also advance in the Artificial Intelligence (AI) algorithms and the Internet of Things. There can be advancements in the building heating systems, like emergence of low-cost, efficient and sustainable heating systems. However, the physical infrastructure changes slow and many traditional heating systems remain in operation likely for many years to come. Where changing of physical infrastructure takes slower steps forward, the new technologies related Big Data and AI are emerging quickly with the help of advances in machine learning, storage capacity and processing power.
To get back to the idea of Big Data, it was previously discussed in this thesis that correlations are more important than causation. In this work, causations have still been in the center of the thinking. Especially the alternative data mining models are based on rethinking and questioning the attributes and their correlations. This
95
rethinking has basis on the understanding of the building’s properties and heating systems. The underlying reasons are dug out to answer the question of why.
However, the alternative models are also based on the values of correlation coefficients. In principle, it would be possible to find the optimal models by just looking at the correlations. Taking a look at possible future trends reveals that the Internet of Things is emerging in the building sector and more sensors are likely to send data related to various conditions room by room. As the amount of data and parameters increases, a thorough understanding of the causation gets more difficult. More accuracy is likely required in forecasting not only the heat demand but also indoor lighting and other indoor climate conditions. In these future cases, it is getting more reasonable to assume the Big Data mindset and let the correlations lead the way.
Berthold et al. (2010) emphasize that human interaction is crucial to make intelligent data analysis project successful, and this is still undeniably true today.
However, the constantly developing Artificial Intelligence and machine learning can possibly take the role of the orchestrator one day: select the optimal input attributes and algorithms to forecast a desired output. What is more, at least one such a solution gets already close today called Ayasdi (Ayasdi, 2015). This software can find the best inputs for a selected output or just find useful insights easily in the data – and at best to replace the need of a trained data analyst.
In the end, the main thing is the accuracy of the forecast, not how much effort we’ve made to make the forecast model and understood more or less important things on the way. Human analysts are biased and select the attributes as an input that are usually of most interest or are thought to be important. AI is steering airplanes and can drive cars, so why not control a building? While waiting for a building to be under the control of an AI, which might not be that far in the future, human endeavor and data mining tools are required in this environment with rapidly increasing sensors and data from the buildings.
96