The rapid advancement of technology-particularly in machine learning and data availability has led to an increasing demand for reducing communication overhead and operational costs. Edge computing is one of the fastest-growing fields, enabling cost reduction and improving communication. Simultaneously, the advancement of machine learning and the availability of data are both accelerating. Federated learning (FL) addresses not only privacy concerns but also issues related to cost and communication eficiency. This paper presents an application of a federated learning framework to improve weather forecast accuracy through the collaborative analysis of distributed data, while ensuring data confidentiality and computational eficiency. To replicate the federated environment, we developed an architecture that combines real-time data collection using the Signal K server, containerization using Docker, and a Hadoop cluster on Microsoft Azure. We evaluated the performance of a Transformer and a Crossformer, demonstrating the efectiveness of both models in this context, with the Crossformer showing superior performance in managing spatiotemporal dependencies for forecasting. Our experimental results indicate an important improvement in reducing the error with respect to previous methods, achieving a Mean Absolute Error (MAE) of 0.144 for the Crossformer and 0.232 for the Transformer, highlighting the potential of FL and advanced deep learning architectures in managing sensitive data in distributed scenarios, in line with previous research trends. This study proposes a robust and scalable approach, opening new perspectives for future applications in cooperative and secure learning.
The exponential growth in the cost for training neural networks has become an increasingly pressing
issue, both in terms of financial cost and computational resources, as highlighted in studies [
One of the most promising approaches to address these challenges is edge computing, which aims
to ofload computation from centralized data centers to distributed local nodes. However, with the
rising amount of data being generated—especially in sensitive domains—the need for secure,
privacypreserving data management becomes equally critical. This is where federated learning (FL) is
relevant [
FL is a recently developed distributed deep learning paradigm in which clients independently train
their local neural network models with private data and then jointly aggregate a global model on the
central server [
The aim of this project is to investigate the practical use of federated learning in collecting and
analyzing multidimensional georeferenced data from dispersed, remote stations. Using cooperative
models will help to improve weather forecasting accuracy by lowering the computing cost to train
the model and maintaining data privacy while also ensuring that, as observed in [
To address this challenge, an architecture that leverages modern technologies such as Signal K server for real-time data collection, Docker for containerization of services, and a Hadoop cluster on Microsoft Azure for federated paradigm simulation has been designed. The main contribution lies in the definition of a complete pipeline, ranging from data collection and pre-processing to the collaborative training phase with advanced federated learning techniques. In particular, the FedProx algorithm was used to put together the weights of diferent models.
By integrating Transformer and Crossformer models for space-time analysis [
The rest of the paper is organized as follows: in Section 2 we discuss the literature; Section 3 illustrates the main architecture, including data collection and data pre-processing; Section 4 discusses the practical implementation of the models; Section 5 details the experimental results of our solution, comparing them to previous works; Section 6 provides a critical interpretation of the results and highlights their implications; finally, Section 7 wraps up the paper’s views and outlines possible directions for future research.
FL was introduced around 2016 by Google as an innovative approach to training machine learning models.
Traditionally, developing high-quality models required access to large amounts of data. This frequently
requires working directly with raw and sensitive data, raising privacy and security concerns [
FL addresses these challenges by allowing model training across multiple decentralized devices,
reducing the need to share or centralize raw data [
Several recent studies address data privacy in FL applied to time-series forecasting with georeferenced sensors. For instance, privacy-preserving collaborative forecasting techniques have been analyzed in energy and time-series domains, demonstrating how data transformation, secure multi-party computation, and diferential privacy can be used to balance forecast accuracy and confidentiality. From a general perspective, data privacy becomes especially important in edge computing scenarios [16].
Reviews focused on IoT sensor aggregation highlight the challenges in resource-constrained settings, such as weather stations. Federated learning approaches combined with secure aggregation have proven efective in residential load forecasting scenarios. These studies support the feasibility of extending federated approaches with diferential privacy or secure aggregation approximations in distributed weather forecasting systems [17].
Based on the previous considerations, privacy is in fact a critical issue for both individual users
and large enterprises, which is why FL has gained significant attention and experienced a boom in
recent years. However, many FL systems still rely on a centralized approach, which can lead to several
problems, such as data falsification and lack of transparency [
While federated learning reduces privacy risks by keeping raw data at the edge, distributed weather forecasting introduces several domain-specific security vulnerabilities. A recent study demonstrates that corrupting a small subset of input observations, even in non-sensitive sensor networks, can significantly skew forecast accuracy. This exemplifies the real risk of poisoning or manipulation at the data collection layer [18].
The study in [
In the realm of renewable energy, Li et al. [21] proposed an approach called federated deep reinforcement learning (FedDRL) for ultra-short-term wind power forecasting. FedDRL integrates the deep deterministic policy gradient (DDPG) algorithm within a federated learning framework not only to improve prediction accuracy but also to enable decentralized model training without compromising data privacy. Climate change impacts, and the dependence on non-renewable energy sources for electricity generation, led to the proposal of numerous IoT-based smart weather station systems [22].
De Vita et al. [
To address the challenges of distributed weather data collection and analysis, we have developed an architecture and methodology that leverages state-of-the-art technologies and distributed machine learning paradigms. The objective is twofold: to ensure the accuracy of forecasts while preserving data privacy and optimizing the use of computational resources. This section describes in detail the key components of our infrastructure, the data acquisition and pre-processing phases, and the adopted federated learning model.
The collection of real-time meteorological data from each station is managed via Signal K Server, an open-source protocol originally designed to aggregate data from marine environments [23, 24].
Federated learning is coordinated by Flower, a framework that manages the aggregation of model weights and provides several algorithms, including Federated Averaging (FedAvg) and Federated Proximal (FedProx). Each of these has specific characteristics, but we chose FedProx to reduce communication overhead between the client and the server, as it employs a quantization process to compress model updates [25, 26]. After local training on each station, the encrypted weights of the models are shared with the coordinating node—never the raw data—thus ensuring privacy. The coordinating node aggregates these weights to create a single global model. Subsequently, updated versions of the global model are sent back to the stations for the next training round, iteratively improving the accuracy of the overall
To collect data from distributed weather stations, we used Signal K, an open-source software protocol originally developed for aggregating sensor data from maritime environments. Signal K was used in this context to gather real-time weather data from sensor-equipped weather stations provided by the University of Naples “Parthenope”. A plugin has been set up in the Signal K system to allow distributed weather stations and a central broker to communicate using the MQTT protocol, allowing time series data to be published and transmitted eficiently over the network. Another plugin has been employed in order to save data locally for model training.
This architecture, illustrated in Figure 1, allowed the central server to receive live environmental measurements from multiple remote nodes with no raw hardware access. The Signal K services were containerized using Docker, ofering a reproducible and scalable deployment environment for data collection across the cluster.
The data input is highly sensitive to data quality. To address this problem, which may lead to imprecise forecasts, we need to do some data pre-processing operations [27].
We created a pre-processing pipeline that includes the following processes in order to standardize and modify this data for our purposes: 1. Wind Direction Encoding: In order to maintain angular relationships and render it suitable for our machine learning assignment, the wind direction was encoded into its sine and cosine, as it is cyclic. 2. Data Smoothing: To reduce noise, each numeric feature was smoothed using a moving average with a window size of 3. Data Aggrega- Hadoop (Azure tion HDInsight) 3. Data Normalization: To increase model stability and make the training process easier, all numerical features-aside from time and direction encoding—were normalized in order to have zero mean and unit variance (i.e., the normal distribution).
In total, there were 20,199 data, collected as follows: 4,196 from Castelvolturno; 6,829 from Città della scienza; 4,919 from Marina di Stabia; 4,255 from Via Acton. The sampling interval was Δ ≈ 1 hour [28].
To motivate the architectural decisions taken during the implementation of our federated learning testbed, we provide a comparative overview of the adopted technologies and their alternatives. Table 1 summarizes the main components and the rationale behind each selection.
In order to process and train the models on distributed data, we used a Hadoop 3.3.4 cluster using Microsoft Azure, as illustrated in Figure 1. The head nodes were configured using the E2 v3 instance type (2 cores, 16 GB of RAM), with two head nodes deployed for redundancy. For fault tolerance and coordination, we also used three zookeeper nodes based on the A2 v2 instance type (2 cores, 4 GB RAM). This central infrastructure provides the coordination and management capabilities necessary for our distributed environment.
The worker nodes were configured using the E2 v3 instance type (2 cores, 16 GB of RAM), with two worker nodes deployed in the cluster. This infrastructure ofers a trade-of between computational eficiency and scalability for our distributed environment. It has been employed to simulate a distributed federated learning environment, where each worker node acted as two diferent clients participating in model training, mimicking automated weather stations collecting and processing meteorological data in a federated manner.
For weather forecasting, we integrated the Transformer and Crossformer models, advanced deep learning architectures suitable for time series analysis with complex dependencies. The Transformer was used for its ability to capture long-range relationships in data sequences [29]. The Crossformer, specifically designed for multivariate and multimodal time series, has been used to handle spatiotemporal interactions between diferent stations [30], but both are based on the Transformer architecture [31].
In addition, both the Transformer and Crossformer architectures had a model dimension of 64, four attention heads, two layers, and a dropout rate of 0.1. The input dimension was 4, including the features station_id, temperature, humidity, and wind_speed, as well as the encoded wind components wind_sin and wind_cos. The output dimension was 1, focusing on temperature. The model operated with an input_window of 20 hours and an output_window of 1 hour. Federated training was conducted over 20 rounds, with 25 local training periods per round.
Model evaluation was conducted to quantify the efectiveness of Transformer and Crossformer architectures in the context of distributed weather forecasting, leveraging the FL approach. The primary objective was to determine the performance of each model in terms of predictive accuracy, taking into account the specificities of the meteorological data collected and the distributed configuration of the training.
The experimental environment has been configured to simulate the real operating conditions of distributed weather stations.
Model performance was evaluated using Mean Square Error (MSE) and Mean Absolute Error (MAE), two of the best metrics for assessing time series predictions [32].
Table 2 compares Transformers and Crossformers models using MSE and MAE metrics. The evaluated
model is the aggregated one obtained at round 20 (the last one) using the FedProx algorithm. Table 3
employs the same approach as the first table; however, it compares our results with the work of De Vita
et al. [
Table 4 shows the performance of the aggregated model (here it refers to the results obtained by the best model between the two, in this case the Crossformer) over the course of the 20 rounds of federated training, showing a progressive improvement in error metrics as the overall model becomes more refined. These preliminary results support the efectiveness of Federated Learning and the architectures chosen for the distributed weather forecasting task.
Figure 2 further compares our approach with that of De Vita et al. [
As can be seen in Table 2, the Crossformer has demonstrated superior performance compared to the Transformer in terms of MSE and MAE (using the fictitious values provided), suggesting a greater ability to manage complex space-dependent weather patterns in meteorological data. This outcome aligns with the Crossformer design, which is tailored for the cross-analysis of multivariate time series [30].
The results obtained, summarized in Table 3, demonstrate an improvement in the performance of
the Transformer and Crossformer models compared to the values reported by De Vita et al. [
As shown in Table 4 and Figure 2, the federated learning setup demonstrates a consistent
improvement in prediction accuracy over the course of the 20 training rounds. Both MAE and MSE decrease
significantly from round 1 to round 20, confirming the efectiveness of the federated learning strategy.
In particular, the MAE drops from 0.510 to 0.144, while the MSE decreases from 0.519 to 0.232, indicating
a more precise model with fewer large errors as training progresses. This improvement is especially
evident after round 10, where the error metrics begin to show a sharper decline. The stability and
convergence observed in the metrics suggest that the federated approach not only preserves data privacy
reduces communication and computational costs but also maintains high predictive performance.
Compared to the baseline values reported in [
While the current testbed consists of a limited number of physical weather stations and a reduced amount of data, this design choice was driven by practical constraints related to data availability and the need to ensure controlled, verifiable experimentation. Given the fragmented nature of data access, we integrated multiple technologies for data collection, selecting solutions that are inherently scalable. In particular, the use of containerized services (Docker), publish-subscribe communication protocols (MQTT), and federated orchestration frameworks (such as Flower) provides a solid foundation for future extensions. Although our experiments are currently limited in scale, the presence of a historical database means that additional data can be incorporated into the architecture to validate performance and scalability at a larger scale. Ultimately, this represents a strategic starting point to explore the real-world applicability of our federated framework in meteorological scenarios.
In this work, we have introduced and developed an innovative application of FL in the context of the collection and analysis of distributed meteorological data. The goal is to improve the accuracy of weather forecasts through collaborative models and to ensure data confidentiality and computing eficiency at peripheral nodes.
We designed and implemented an architecture that leverages modern technologies such as Signal K for real-time data collection, Docker for service containerization, and a Hadoop cluster on Microsoft Azure for federated paradigm simulation. A key contribution of this work is the definition of a complete pipeline that covers data collection and preprocessing up to the collaborative training phase, using advanced FL techniques.
The integration of Transformer and Crossformer models for space-time analysis has demonstrated the efectiveness of FL techniques in real scenarios, overcoming the limitations of data centralization, and paving the way to new safe and cooperative learning scenarios. In particular, it was highlighted that the Crossformer has outperformed the Transformer in space-time data management for weather forecasting. The experimental results show that the Transformer model had an MSE of 0.265 and an MAE of 0.194, while the Crossformer model had an MSE of 0.232 and an MAE of 0.144.
However, an important consideration is about data security, integrity, and reliability. Attacks could
potentially originate not only from individual nodes but also from the central coordinating node [
Data were collected from distributed meteorological stations of the University of Naples “Parthenope”, with a total of 20,199 data from diferent locations. To ensure data quality and improve the stability of the model, a pre-processing phase has been implemented which includes wind direction coding, smoothing of data, and normalization of numerical features.
In summary, this study not only confirms the efectiveness of Federated Learning and deep learning models such as Transformer and Crossformer in weather forecasting but also proposes a robust and scalable approach for managing sensitive data in distributed environments, opening new perspectives for future applications in cooperative and safe learning environments. Edge computing has proven to be an instrumental tool in optimizing resource utilization, leading to significant economic and communication eficiencies.
It could be interesting to explore possible directions for future work. An interesting path could be
starting from the work of De Vita et al. [
The author(s) have not employed any Generative AI tools. [15] S. Shen, T. Zhu, D. Wu, W. Wang, W. Zhou, From distributed machine learning to federated learning: In the view of data privacy and security, Concurrency and Computation: Practice and Experience 34 (2022) e6002. [16] C. Gonçalves, R. J. Bessa, P. Pinson, A critical overview of privacy-preserving approaches for collaborative forecasting, International journal of Forecasting 37 (2021) 322–342. [17] J. D. Fernández, S. P. Menci, C. M. Lee, A. Rieger, G. Fridgen, Privacy-preserving federated learning for residential short-term load forecasting, Applied energy 326 (2022) 119915. [18] E. Imgrund, T. Eisenhofer, K. Rieck, Adversarial observations in weather forecasting, arXiv preprint arXiv:2504.15942 (2025). [19] S. Chen, G. Long, T. Shen, J. Jiang, C. Zhang, Federated prompt learning for weather foundation models on devices, arXiv preprint arXiv:2305.14244 (2023). [20] P. Hamedi, R. Razavi-Far, E. Hallaji, Federated continual learning: Concepts, challenges, and solutions, arXiv preprint arXiv:2502.07059 (2025). [21] Y. Li, R. Wang, Y. Li, M. Zhang, C. Long, Wind power forecasting considering data privacy protection: A federated deep reinforcement learning approach, Applied Energy 329 (2023) 120291. [22] S. Ganesan, C. P. Lean, L. Chen, K. F. Yuan, N. P. Kiat, M. R. B. Khan, Iot-enabled smart weather stations: innovations, challenges, and future directions, Malaysian Journal of Science and Advanced Technology (2024) 180–190. [23] D. Di Luccio, A. Riccio, A. Galletti, G. Laccetti, M. Lapegna, L. Marcellino, S. Kosta, R. Montella, Coastal marine data crowdsourcing using the internet of floating things: Improving the results of a water quality model, IEEE Access 8 (2020) 101209–101223. [24] R. Montella, D. Di Luccio, L. Marcellino, A. Galletti, S. Kosta, G. Giunta, I. Foster, Workflowbased automatic processing for internet of floating things crowdsourced data, Future generation computer systems 94 (2019) 103–119. [25] T. Li, A. K. Sahu, M. Zaheer, M. Sanjabi, A. Talwalkar, V. Smith, Federated optimization in heterogeneous networks, Proceedings of Machine learning and systems 2 (2020) 429–450. [26] S. Dasari, R. Kaluri, 2p3fl: A novel approach for privacy preserving in financial sectors using lfower federated learning, CMES-Computer Modeling in Engineering and Sciences 140 (2024) 2035–2051. [27] L. Zheng, W. Hu, Y. Min, Raw wind data preprocessing: a data-mining approach, IEEE Transactions on Sustainable Energy 6 (2014) 11–19. [28] S. Wang, J. Cao, S. Y. Philip, Deep learning for spatio-temporal data mining: A survey, IEEE transactions on knowledge and data engineering 34 (2020) 3681–3700. [29] H. Zhou, S. Zhang, J. Peng, S. Zhang, J. Li, H. Xiong, W. Zhang, Informer: Beyond eficient transformer for long sequence time-series forecasting, in: Proceedings of the AAAI conference on artificial intelligence, volume 35, 2021, pp. 11106–11115. [30] Y. Zhang, J. Yan, Crossformer: Transformer utilizing cross-dimension dependency for multivariate time series forecasting, in: The eleventh international conference on learning representations, 2023. [31] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, I. Polosukhin,
Attention is all you need, Advances in neural information processing systems 30 (2017). [32] A. Botchkarev, Performance metrics (error measures) in machine learning regression, forecasting and prognostics: Properties and typology, arXiv preprint arXiv:1809.03006 (2018).