Smart homes have become increasingly popular with the widespread adoption of Internet of Things (IoT) devices ^{[1][2][3][4][5]}[1,2,3,4,5], and have paved the way for improving multiple aspects of homes by utilizing the enormous amounts of data generated every day. One key challenge in this domain is predicting energy consumption to optimize energy management, reduce waste, and save costs ^[6][7][8][6,7,8]. Several studies have investigated this problem, including both centralized and decentralized approaches ^[9][10][9,10].

Traditional centralized prediction models, such as regression and time-series analysis, require data to be collected and processed on a central server ^[11][12][13][11,12,13]. However, collecting and transmitting sensitive data from smart homes to a central server can pose privacy concerns. Additionally, these methods do not scale well to large datasets, and the central server can become a bottleneck in the prediction process.

To address these challenges, decentralized approaches, such as FedAvg from federated learning (FL), have emerged [14]. FL enables multiple clients to train a machine learning model collaboratively without sharing their raw data. FL has been successfully applied to energy load prediction for smart homes, improving prediction accuracy while preserving data privacy ^[15][16][17][15,16,17]. Current FL approaches do not take into consideration the nature and additional properties of data. More specifically, limited work has been conducted in the area of federated learning regarding time-series datasets ^[18][19][18,19]. Nevertheless, none of them take into consideration and exploit the property of age.

2. Federated Learning-Based Consumption Prediction in Smart Homes via Age-Based Model Weighting

One of the early works on smart homes consumption prediction was conducted by [20], where they presented an energy management system (EMS) for smart homes. This system uses a data-acquisition module, which is an IoT device with a unique IP address, to interface with each home device. This creates a mesh wireless network of devices. The module, called the system on chip (SoC), collects energy-consumption data from each smart home device and sends them to a central server for analysis. The energy consumption data from all residential areas are collected in the utility company’s server, which results in a large collection of big data. The proposed related work makes use of standard business intelligence (BI) and big data analytics software to manage energy consumption effectively and meet consumer demand. More recently, deep learning techniques have been applied to multiple domains, for example, the area of manufacturing ^[21][22][23][21,22,23]. In the area of smart home consumption prediction, ref. [24] proposed a convolutional neural network (CNN) based model for predicting the electricity consumption of smart homes. Similarly, ref. [25] proposed a long short-term memory (LSTM) based model for predicting the energy consumption of smart homes. They experimented on multiple datasets showing the effectiveness of LSTMs. Federated learning (FL), also referred to as collaborative learning, is a machine learning method that enables training an algorithm without transferring data samples between various decentralized edge devices or servers that store local data samples. This approach distinguishes itself from conventional centralized machine learning techniques, where all local datasets are uploaded to a central server, as well as traditional decentralized alternatives that often assume a uniform distribution of local data samples. Numerous studies have investigated the application of FL in predicting energy consumption in smart homes. Previous work [26] suggested two approaches to decrease the costs associated with uplink communication. The first approach involves utilizing structured updates, which involves learning an update from a limited parameter space that is represented by a smaller set of variables. This can be achieved through techniques like low-rank approximation or applying a random mask. The second approach, known as sketched updates, entails learning a complete model update and then compressing it using a combination of quantization, random rotations, and subsampling before transmitting it to the server. Experimental results on convolutional and recurrent networks demonstrate that these proposed methods can reduce communication costs. Ref. [14] introduced a practical technique for federated learning of deep networks using iterative model averaging. They conducted a thorough empirical assessment, utilizing five distinct model architectures and four datasets. The results of these experiments indicate that the proposed approach remains resilient even when confronted with unbalanced and non-independent and identically distributed (non-IID) data distributions, which are common characteristics of this scenario. The authors focused on reducing communication costs, which are a primary constraint in federated learning. They demonstrated that their method significantly reduces the number of communication rounds required, achieving a reduction of 10–100 times compared to synchronized stochastic gradient descent. Ref. [27] presented a system that allows for training a deep neural network using Tensor-Flow on data stored on a mobile phone. The data remain on the device and are not shared. The weights are combined in the cloud using federated averaging, creating a global model that is sent back to the phones for inference. To ensure privacy, secure aggregation is used to make sure individual updates from phones are not viewable on a global level. This system has been used in large-scale applications, such as phone keyboards. The approach addresses several practical issues, including device availability, which depends on the local data distribution in complex ways, unreliable device connectivity, interrupted execution, coordinating execution across devices with different availability, and limited device storage and computing resources. In recent work by [18], a federated series forecasting framework was proposed by redesigning a hybrid model that enables neural networks, utilizing the extra information from the time series to achieve time-series-specific learning via exponential smoothing. Regarding smart homes and energy prediction, a number of approaches have been introduced. In their position paper, ref. [15] proposed a novel architecture for smart homes, called IOTFLA, focusing on the security and privacy aspects, which combines federated learning with secure data aggregation. Ref. [17] proposed a prediction model based on the analysis of the energy usage patterns of the households. They used a clustering algorithm to group the households with similar energy consumption patterns and then trained a prediction model for each cluster. The results showed that their model can accurately predict the energy consumption of households. Moreover, the non-independent and identically distributed (non-IID) data samples across participating nodes slow model training and impose additional communication rounds for FL to converge. Ref. [28] proposed the federated adaptive weighting (FedAdp) algorithm that aims to accelerate model convergence under the presence of nodes with the non-IID dataset. In the work [29], the authors employed privacy-preserving principal component analysis (PCA) to extract features from data obtained from smart meters. Using this approach, they trained an artificial neural network in a federated manner, incorporating three weighted averaging strategies. The goal was to establish a connection between the smart meter data and the socio-demographic attributes of consumers. Ref. [30] suggested a personalized federated learning (PFL) based user-level load-forecasting system. Using local data, the derived personalized model performs better than the global model. To add another layer of privacy protection to the suggested system, the authors also used a unique differential privacy (DP) method. Based on the generative adversarial network (GAN) theory, the method balances prediction accuracy and privacy throughout the game. By performing simulation tests on real-world datasets, they demonstrate that the proposed system can meet the requirements for accuracy and privacy in practical load-forecasting scenarios. To address the long-term optimization considerations for latency, accuracy, and energy consumption in wireless federated learning, ref. [31] introduced a mixed-integer optimization problem. The objective was to minimize the cost function over a finite number of rounds while adhering to the energy budget constraints of each client in the long run. To tackle this optimization problem, the authors proposed an online algorithm called per-round energy drift plus cost (PEDPC), which consists of two main components: client selection and bandwidth allocation. The client selection is addressed using the increasing time-maximum client selection (ITMCS) algorithm, while the barrier method is employed for bandwidth allocation. This approach allows for effectively handling the optimization problem in a real-time fashion. Advances and open problems as well as future directions in federated learning have been described in recent papers by ^[27][32][33][27,32,33]. While these studies have shown the promise of FL for energy-consumption prediction in smart homes, there is still room for improvement. In particular, the use of FL for prediction models that can be implemented on resource-constrained smart home devices remains a challenging problem. None of the aforementioned methods exploit the age of datasets within the clients.