2.1 Predictive Machine Learning for Energy consumption

Research done using a data driven approach for predicting energy ratings focusing on reducing dimensions has benefited field the most. Balanced accuracy along with model simplicity was achieved by [22]. The accuracy was moderate however the complexity was reduced to quite an extent. The novel model QLattice was implemented to achieve high accuracy with the main aim of striking transparency and accountability of decisions taken by the algorithms. This methodology was proposed to be integrated with Explainable AI to enable the end user understand and control the decision making process. By reducing the input dimensions and making the model simpler to use, the research reinstated that ML is efficiently used for predictive problems using lesser features. [28] also developed a regression model implementing ANN by simplifying the input features for users. The authors used only 12 features for energy rating prediction and thus made the process usable by the end user with less domain knowledge.Some studies have also been done by using synthetically generated data for building stock. [19], highlighted in their research the need for substantial quality data to understand the role of various parameters in predicting the energy rating. [20], formulated synthetic data generator that uses only 19 input features. The issues with finding the right data can be resolved by using synthetic data for energy performance analysis. However, this data cannot be validated with the real world scenarios but it is useful for carrying out any research in the similar field The researchers used J and energy tools for generating synthetic building stock.Other similar studies done by [23], [29] and [30], translate the efficiency of ML models in predicting energy ratings with reduced features than used for actual derivation of the rating. The first author applied statistical techniques on building stock data for Estonia and focused on understanding zero energy utilisation for domestic buildings We will stop the model achieved an accuracy of 73 percent but it was mostly dependent on the historical energy performance certificate data of European countries. Miller, however, used explainable machine learning and focused mainly on non residential smart meter data for classification of energy ratings. The author used temporal feature engineering for identifying the key features for breeding classification.An interesting insight provided by a [27] into the behavioural impact on energy rating by occupants Helped in extending the use of machine learning models predicting energy ratings. The author also analysed and provided criticism on the assumptions made by DEAP Software used for the calculation of energy ratings. Another different use case for machine learning models in predicting energy ratings for detection of data corruption in energy ratings are calculated using default values of building properties the authors [25] detected abnormalities in the data for similar features of the dwelling They were awarded different energy ratings. This study was performed on Irish BER data and questioned the data integrity of BER assessment process.

2.2 Optimization for Retrofit Analysis

Studies that extend the use of machine learning models for prediction of energy ratings into understanding and analysing retrofits solutions for domestic buildings Help urban planners build energy efficient homes and reduce carbon emissions. The [31] combines machine learning model with multi objective optimization in a large data set. The study demonstrates significant energy savings and return on investments. The objectors formulated in this study include minimising the energy consumption of a building and maximising the cost savings for retrofit solutions based on building parameters that could be modified in an existing building. The [32] use a novel combination of ELM and shell game optimization focusing on smart metering data collected from European countries The machine learning models Used in the research like ANN provide moderate accuracy as the main objective of lies in understanding the impact of retrofit solutions on energy efficiency.The [26] estimates the effect on property sale and rental prices based on the energy rating of the dwelling. The study uses mathematical and statistical tools to determine the effect of energy efficient homes on property values. [17], Discusses the inter-building energy dynamics and inter-dependencies of adjacent buildings on the energy efficiency. The research focuses on a data driven machine learning model that uses simulation to understand the energy consumption of single building, community building urban and rural buildings providing valuable insights for building design stage and policy formulation for urban energy efficiency for a sustainable future.

3.1 Data Collection

Analysing any data and gathering insights from it largely depends on how the data was collected, organised and stored for the process. The Building Stock data for Ireland published by Sustainable Energy Authority of Ireland (SEAI) 44 This data is published every quarter and is a part of the BER Research Tool available for public use. The essence of the dataset is tabulated in Table 1. Along with these columns, there are in total 211 features in the dataset. The research focuses only on building stock for county Dublin that gives substantial data of 82 thousand dwellings for analysis and prediction. The data used has values for building stock from years 1700 to 2024.The data used for this research has 82,030 rows and 211 columns. The data can be divided into:

As mentioned in Table 1 these features are used to mathematically calculate the BER for each dwelling using DEAP(Dwelling Energy Assessment Procedure) software55 This research aims to predict the value of BER based on the features of building stock using a data driven approach.

3.2 Data Cleaning and Pre-Processing

Data Cleaning stage comprises of understanding the quality of data and its form based on planned implementation. It involves checking for inconsistency in the dataset. Checking if there are missing or invalid values in the data helps in making the data ready for analysis. The dataset used was large and mostly clean. However, there were some columns that had missing values or null values. Some columns were not relevant to this study. Data cleaning was performed in two ways:

After removing the features using domain knowledge, 161 features were analysed for missing values. On dropping features with more than 50 percent missing values, 121 features remained. They were further analysed and visualised to take appropriate actions.

3.3 Data Visualisation and Transformation

The Data is not used as it is for Machine Learning and processing. The dataset includes features with varied data types and values. For dimensionality reduction, there are various techniques that select important features that impact the Energy Rating most. Feature Importance analysis like Random Forest, Decision Tree Classifier are applied to decide the relevant features in the dataset. This process has a major contribution to the accuracy of the model and its prediction. For understanding variance in the data, PCA or Principal Component Analysis is performed. The dataset is reduced to its first two principal components (PC1 and PC2) that are independent of each other.

Figure 5 PCA plot in feature space(original illustration).

Figure 5 shows the scattered distribution of different BERs for each dwelling in the dataset against the Principal components identified by PCA. There seems to be a noticeable separation between the different energy ratings, like between 'A' and 'E/F/G'. This separation suggests that the PCA has successfully captured the differences in the dataset's features and it distinguishes high energy ratings (like 'A') from lower ones (like 'E/F/G'). The overlap between the ratings, especially between 'B' and 'C', or 'C' and 'D'. It suggests that there is some similarity in the features of the building stock. The plot shows that the ratings are not linearly separable and overlap in the feature space.

3.4 Modelling and Experimentation

The aim of the research is to compute a machine learning model that predicts the BER for a dwelling based on the features present in the dataset. In this study, several models are applied to predict and classify the Energy ratings of the dwellings based on the features present in the data set. This study adopts the sample technique for data splitting by dividing the training and test data sets in a ratio of 80 and 20 with a random state of 42. There are many various techniques that could be used for dividing the data set into these two subsets to ensure that the division is made purely random. Many researchers like [20] use cross validation or K Fold for data splitting. As the focus of this research is on studying the impact of the retrofit solutions and not on the basic modelling and hence this data splitting is suitable for the research.This research uses LazyPredict library of Python that applies 25 Machine Learning Classifier Models and concludes with the accuracy and other metrics for each model. This helps in identifying the best model suitable for the dataset in hand. Figure 6 shows the metrics of all the models implemented on the dataset. LightGBM (Light Gradient Boosting Machine) classifier turned out to be the best performing model in terms of accuracy and F1 Scores.

Figure 6 Model comparisions(original illustration).

CRISP-DM is an iterative process, where modifications can be made to check any improvements in the result. Four combinations were performed to check the best conditions that fulfill both research objectives. As this research focused on the retrofit solutions as well, feature selection was an important step in order to achieve desired impact on the energy rating. The Table 2 shows that even though the accuracy obtained while using 132 features was the maximum at 98 percent, implementation of experiments would have been affected. The feature selection process in this iteration did not include domain knowledge which would be useful in identifying features that are related to experimental features. After performing all possible scenarios, modelling was done by selecting features using Domain knowledge based on experiments to be performed and dropping columns that 50 percent empty.

The second part of this research aims to experiment retrofit changes in existing dwelling features and analyse the impact of these changes. If these changes help in improving the energy ratings, these could be used by dwellers to improve the energy efficiency of their dwellings. Following are the experiments made on the Test Dataset:

3.5 Evaluation

The various models applied to the data set are then compared with each other based on the accuracies, balance accuracies, ROC AUC, F1 Score and the time taken. The accuracy calculated for each model is the difference between the predicted value calculated by the model and the actual truth value of the test data set.The formula for F1 Score is:

where:

The formula for Accuracy is:

In this research, retrofit solutions are implemented as experiments. Features such as heat pumps, solar water heat pumps, heat control systems, etc, are modified independently as different experiments. The machine learning model selected as the best performing model is then used to predict the energy rating of these individual experiments. The ratings predicted by the model are then compared with each other to see if they are significantly different from each other. If the predicted values for each energy rating are significantly different for each experiment then it can be concluded that the changes or the retrofit solutions made to the data set impacted the rating positively. To evaluate the performance of the experiments and to test the hypothesis one way ANNOVA and T tests are used. The F statistics and the P values are then analysed to conclude the findings.

4.1 Experiment 1

Installing Heat Pumps in dwellings with no Heat Pumps was a retrofit that should improve the energy rating of the building. However, the count of dwellings whose energy rating improved was just 1 as seen in Table 4. For other impacted dwellings, the rating either did not improve or it was degraded to a lower rating. This however could have been due to other factors like dependency of Heat Pump installation on other features in the dataset.

4.2 Experiment 2

Installing Solar Hot Water system in dwellings with no Heat Pumps was a retrofit that should improve the energy rating of the building. However, the count of dwellings whose energy rating improved was 29 as seen in Table 5. For other impacted dwellings, the rating either did not improve or it was degraded to a lower rating. This however could have been due to other factors like dependency of Solar Hot Water system installation on other features in the dataset.

4.3 Experiment 3

Enabling Central space heating Boilers fueled by Gas with a room thermostat, saves energy consumption by regulating the heat based on the room temperature. This should improve the energy rating of the dwelling. However, the count of dwellings whose energy rating improved was just 1 as seen in Table 6. For other impacted dwellings, the rating either did not improve or it was degraded to a lower rating. This however could have been due to other factors like dependency of Boilers on other features in the dataset.

4.4 Experiment 4

Enabling Central space heating Boilers fueled by Oil with a room thermostat, saves energy consumption by regulating the heat based on the room temperature. This should improve the energy rating of the dwelling. However, the count of dwellings whose energy rating improved was just 2 as seen in Table 7. For other impacted dwellings, the rating either did not improve or it was degraded to a lower rating. This however could have been due to other factors like dependency of Oil fueled Boilers on other features in the dataset.

4.5 Discussion

The research was able to achieve both the research objectives. However, the point of discussion is made around what was done to achieve those objectives and at what was compromised in achieving them. Few arguments that include constructive criticism towards the research are discussed in this section.

The proposed Hypothesis in this research were evaluated by using Statistic inferences.

Figure 7 Retrofit analysis visualisation (original illustration).

To get statistical insights One Way ANNOVA was performed which obtained a value of 0.19 for F-Statistics and a P value of 0.9. This was performed between the predicted energy ratings for all 4 experiments and the original energy rating for Test Data. The T tests performed for statistical inferences are discussed below:

The T Tests and ANNOVA, conclude that there is no significant difference between the mean of all 4 experiments and the Predicted test value. It means that the research failed to reject the Null Hypothesis and that the Retrofit solutions did not significantly impact the energy rating of the dwellings. It can be confirmed through data by observing the count of dwellings with improved energy ratings as shown in Figure 7.

It would have been an ideal case if the suggested retrofit solutions significantly improved the energy ratings of dwellings, however, the research answered the extent to which Machine learning data driven approach can help in predicting the energy efficiency of domestic buildings.