Privacy preserved and decentralized thermal comfort prediction model for smart buildings using federated learning

Contribution

The research’s primary contribution is presented in the list format below.

• This article proposed an approach for the thermal comfort model that enhances predictions of thermal experience by concentrating on important features using the ASHRAE RP-884 standard dataset. The novelty lies in the privacy preservation of the individual’s data and improved performance using federated learning.

• The proposed approach first builds the local model and then delivers the variables of the global model. Subsequently, the global model integrates the revised variables and enhances the process while preserving the privacy of every client’s information.

• The proposed method employs the min-max normalization technique to normalize the data and leverages SMOTE to enhance the depiction of the minority class’s perception. The global model receives its parameters from the local model, which is trained initially through the FDL paradigm. Next, the global model ensures that each client’s data is private by aggregating the new parameters and training the system.

• The experiment demonstrates that the FDL model is based on a deep neural network, reaching 82.40% accuracy while offering client data protection.

Organization

The following sections comprise the article: The heat balancing and ML and FL methodologies for analyzing the thermal comfort model are discussed in ‘Related Work’. ‘Federated Deep Learning Framework’ explains the proposed work’s research methodology using the ASHRAE RP-884 standard dataset and DNN model. The results are explained and discussed in ‘Experimental Results and Analysis’. ‘Conclusion and Future Scope’ contains the study’s summary and recommendations for further research.

Experimental dataset

The data is gathered through field surveys covering 160 different building locations worldwide. It is offered as part of the ASHRAE project to establish a thermal comfort preference model. As part of the ASHRAE RP-884 public repository, ASHRAE manages a collection of files from multiple studies conducted by different researchers. Data files are collected from several environmental zones dispersed throughout different geographical areas (De Dear & Schiller Brager, 2001). The dataset was chosen because thermal comfort is crucial for human health and productivity. A lack of thermal comfort may cause inconvenience among building occupants. The ASHRAE RP-884 dataset utilized in this study has 12,595 records and 56 characteristics. The creation of an adaptive model is the goal of this dataset. It comprises over 20,000 customer comfort votes from 52 studies conducted across 10 different climatic zones. The dataset contains 55 primary identifiers such as blcode, sub, ash, ACT10, ACT20, ACT30, ACT60, met, clo upholst, insul, and TA_H with three classes: UW, N, and UC. There are 20,757 training record totals. Because the present investigation works with two clients, each with unique data collection, the dataset is split into two windows. Figure 2 presents the overall geographic distribution of building studies comprising the RP-884 thermal comfort database.

Data preprocessing

Data preprocessing in ML refers to transforming raw information into a format that can be used to develop and improve ML models. Data preprocessing is the first step in ML, preceding the construction of a model. Data preparation is essential to increase data reliability and extract useful information. Real data often lacks specific attributes or patterns, is partial, erroneous (has mistakes or anomalies), and cannot be trusted. Data preparation is crucial in this case because it makes it simpler to arrange, filter, and present raw data in a way that ML models can use. To balance and normalize the data, the dataset is evaluated using two preprocessing techniques: SMOTE and normalization. Using imitation samples to depict the minority classes, SMOTE is an oversampling technique. The overfitting issue from random oversampling is mitigated partially using this strategy.

Min-max normalization: Min-max scaling normalizes the features, retains data integrity, and reduces variation in the uncertain dataset. Models depending on value amplitude need to scale the input attributes. For this reason, the discrete interval of real-valued numerical qualities between 0 and 1 is defined by normalizing. To normalize the data, Eq. (3) is employed. (3)${\displaystyle Y_{norm}=\frac{Y_i-Y_{min}}{Y_{max}-Y_{min}}.}$

The dataset is cleaned up and then split into training and testing halves, with 25% of the dataset destined for testing and 75% for training.

Model architecture

This study looked into DNN’s potential for estimating the power consumption of buildings. The DL technique includes the area of learning that uses nonlinear data over several phases through organizational structures, as Dong and Deng mentioned earlier (Deng & Yu, 2014). Deep learning is usually used chiefly for pattern recognition and learning. The lower-level elements in the hierarchy of deep architecture are used to build higher-level principles. DL combines NN, pattern identification, and visual design. The DL model forecasts satisfactorily with large data sets. According to Huang et al. (2013), when compared to other ML classifiers, DNNs are extraordinary. The proposed DL model examines the inherent trends in the heating and ventilation load of a structure.

Further, multi-task learning, which considers the attributes of all construction forms using a single-layer deep neural network, makes sense in DL. A DNN is an array of hidden layers in a feed-forward NN. The primary modules in this network are multi-layered self-learning modules. A DNN divides its input and output layers using hidden units. The scalar variable y_p of the next layers can be translated by the hidden unit p using a logistic function to the input x_p beneath. In a DNN network, ith neuron x_i’s output can be estimated using Eqs. (4) and (5): (4)${\displaystyle x_i=f\left({\xi }_i\right)}$ (5)${\displaystyle f\left({\xi }_i\right)=\vartheta +{\Sigma }_{j\varepsilon {\tau }_i}-1W_i{X}_j.}$

The ith-neuron’s potential is denoted by ξ_i, whereas the transfer function is represented by f(ξ_i). The following Eq. (6) depicts the transfer function: (6)${\displaystyle f\left({\xi }_i\right)=\frac{1}{1+exp\left(-{\xi }_i\right)}.}$

The goal cost function where the target values, y_o and $\widehat{y_o}$, is determined by the output neurons can be represented as the sum of squared errors as shown in Eq. (7). (7)${\displaystyle C=\Sigma 1/2\left(y_o-\widehat{y_o}\right).}$

This study used a single-input layer sequential DNN approach. Using 256 units and the relu activation function, the input layer has the shape of a 56. Next, we have the hidden layer, which incorporates four dense layers. Relu is the activation function, and the units that incorporate the four dense layers are 128, 64, 32, and 28. The softmax function is used in the output layer, which comes next and has three units. Each dense layer employs a fully linked layer, and the activation functions softmax and relu to solve the classification issue. The DNN model has employed Adam as an optimizer to compute and minimize loss using sparse_categorical_crossentropy. The model size is 64 kb. Figure 3 presents the DNN model’s framework.

Server-based training using log data

The FL paradigm uses three primary parameters: one central server and two clients. The FL server directs client selection and collects incoming model improvements during the model training process. Server-logged records are used to train the DNN framework. Logs are confidential and anonymous before the training. The experiment is run three times as though there are three rounds, with various parameters initially for the federated learning model’s server setup. Every round of the FL model consists of two phases: the assessment cycle and the fit round. In the fit round, clients send their training results to the server; both provide their testing results in the evaluation round, and the server aggregates the data. The server evaluates records from N clients and determines that the DNN method has an accuracy of 82.40%.

Federated training results of client 1

After each round, Table 1 shows an improvement over the previous round. In the first round, 76% accuracy is achieved, which increased to 80% and 85% in rounds 2 and 3, respectively. Similarly, precision is obtained at 76%, 82%, and 86% in rounds 1, 2, and 3, respectively. The recall value is achieved at 85%, 87%, and 90%, and the f1-score is 80%, 84%, and 88%.

Figure 4A showing training and validation accuracy, Fig. 4B loss, and Fig. 4C ROC of round 3. In Fig. 4A, the training and validation accuracy curve is represented. The lowest training accuracy is 0.76% a, reaching a maximum of 0.89% accuracy. The minimum validation accuracy is 0.77%, which increased to 0.85%. Figure 4B graphically represents the training and validation loss graph. The training loss started from 0.55% and decreased to 0.26 and the validation loss started from 0.52% and lessened to 0.40%.

Figure 4C shows the class’s receiver operating characteristic (ROC) curve. Three classes were used in the experiment where the ROC curve of class 0 is represented by a blue line with an area of 0.92, the ROC curve of classes 1 and 2 is represented by an orange line with an area of 0.98, and the green line with an area of 0.97 respectively by showing best outcomes on the used dataset for the proposed framework. The ROC curves in the leftmost position indicate a performance improvement. The ROC curves in the upper-left corner show enhanced efficiency.

Federated training results of client 2

Client 2 performance is less than that of client 1, shown in Table 2, and achieved an accuracy of 72% in the first round, which increased to 80% and 83% after the second and third rounds. Similarly, precision is achieved at 72% in round 1, 80% in round 2, and 84% in round 3. The recall and f1-score are 76% and 75% in round 1, 88% and 84% in round 2, and 89% and 87% in round 3, respectively.

Figure 5 graphically represents the training and validation accuracy curve, loss curve, and ROC curve of round 3 of client 2. Figure 5A indicates the training and validation accuracy curve. The lowest training accuracy is 0.76% following some oscillation between losses and gains, comes 0.90%. The validation accuracy is 0.74% then, after some fluctuation, it increases to 0.83%. The training and validation loss curve is represented in Fig. 5B. the training loss begin from 0.55% and lessen to 0.25. The minimum validation loss is 0.57% and diminishes to 0.45%.

Figure 5C shows the class’s ROC curve. Three classes were utilized in the experiment where the ROC curve of class 0 is represented by a blue line with an area of 0.90, the ROC curve of classes 1 and 2 is represented by an orange line with an area of 0.97, and the green line with an area of 0.96 respectively by displaying best results on the used dataset for the proposed approach. The ROC curves in the leftmost position indicate a performance improvement. The ROC curves in the upper-left corner show enhanced efficiency.

The proposed methodology is more successful with fewer false positive and false negative statistics and more persistent, improved real positive and negative results. Last, the dataset can be correctly categorized using the current methods. The confusion matrix of the proposed methodology is shown in Fig. 6A, which also provides a performance summary for a classification algorithm. Figure 6B illustrates the proposed methodology’s confusion matrix and summarises the classification algorithm’s performance so that the dataset can be correctly categorized using the current methods.

Comparison analysis

The study compares with the existing study (Zhang & Li, 2023) as shown in Table 3. According to Zhang & Li (2023), the accuracy is 62.6% accuracy, 57.0% precision, and 59.0% F-score, whereas the proposed method obtained 85.0% accuracy, 86.0% precision, 90.0% recall and 88.0% F-score. The proposed technique performs better in this context than the Base Paper, indicating that our technique outperforms in thermal comfort model prediction for smart buildings.

Discussion and analysis

An essential component of smart buildings that helps with the development, analysis, and realization of intelligent structures is thermal comfort. These smart buildings’ energy usage forecasts are essential since resource efficiency involves complex decision-making procedures. To do the statistical analysis, the accuracy, precision, recall, and F1-measure optimization metrics are used to assess the performance of the proposed model. Results from federated learning models are assessed for performance, significance, and potential for generalization through statistical analysis. Federated learning proposes a distinctive set of challenges compared to traditional centralized ML methods. It raises issues related to delays in communication and bandwidth limits by decentralizing model training among devices. It necessitates an immense amount of resources to transmit model updates from the devices to the central server. It is frequently applied in situations when protecting user privacy is crucial. Complexity arises from ensuring private data stays on local devices while adding to the global model. Securing the model updates and communication channels is essential since decentralized learning can be more susceptible to attacks. In a federated learning environment, devices may differ regarding data distribution, network circumstances, and hardware capacities. This heterogeneity can make the process of upgrading and aggregation more difficult. The central server must compile model updates from various devices. It is difficult to choose an appropriate aggregation approach that combines the inputs from various devices. To overcome all these challenges, we employed the federated deep learning model with various approaches, Including penalty terms to the loss function, and regularisation can lower model complexity. The recommended FDL framework ensures adherence to privacy regulations by implementing responsible and open data processing practices and incorporating privacy-protecting elements. A federated learning-based deep learning model increases prediction performance. This work uses the FDL model to predict the thermal environment for intelligent constructions. The experiment’s outcomes demonstrate that the proposed FDL model outperforms traditional machine learning methods regarding accuracy and efficiency.