⚡ electrical engineering

An Adaptive Decentralised Federated Learning Framework for Load Forecasting in Dynamic Microgrid Environments

This paper proposes a secure and scalable Decentralised Federated Learning framework integrated with a hybrid TCN-BiLSTM architecture to achieve high-accuracy load forecasting in dynamic microgrids while overcoming the privacy, reliability, and resource constraints of traditional centralized approaches.

Original authors: Md. Julkar Nayeen Mahi, Ranesh Naha, Alistair Barros

Published 2026-09-02

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Md. Julkar Nayeen Mahi, Ranesh Naha, Alistair Barros

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the modern world, electricity is no longer a one-way street flowing from a massive power plant to a passive home. Instead, neighborhoods are becoming their own mini-power grids, known as microgrids. These systems mix traditional electricity with local sources like rooftop solar panels and battery storage, allowing homes to generate and share power with their neighbors. To keep these delicate networks stable, a smart computer system called an Energy Management System must constantly guess how much power will be needed in the coming hours. If it guesses wrong, the grid could become unstable or waste expensive energy. The challenge is that every home uses electricity differently; some families run air conditioners all day, while others only turn on lights in the evening. Furthermore, the sun does not shine with perfect consistency. To make accurate predictions, the system needs to learn from the patterns of thousands of different households. However, asking every home to send its detailed energy usage data to a central computer creates a serious privacy risk, as these records can reveal exactly when people are home, what appliances they use, and their daily routines.

Researchers at the Queensland University of Technology have developed a new way to solve this puzzle without ever asking a home to reveal its private secrets. They created a system where the learning happens locally, right on the smart meters inside each house, and only the "lessons learned" are shared with the neighbors. Imagine a group of students trying to solve a difficult math problem. Instead of handing their notebooks to a teacher who grades them all at a central desk, each student works on their own. When they finish, they simply tell their neighbors what the answer key looks like, without showing their own work. The neighbors then combine these hints to improve their own understanding. This approach, known as decentralised federated learning, allows the entire network to become smarter about predicting energy use while ensuring that no single computer ever sees the raw data of a specific household.

The researchers found that simply sharing these hints was not enough to handle the wild swings in energy demand caused by weather changes or sudden spikes in usage. They discovered that standard learning tools often forgot the long-term trends or missed the quick, sharp changes in power consumption. To fix this, they built a hybrid learning engine that combines two different types of artificial intelligence. The first part acts like a high-speed camera, capturing the rapid, second-by-second fluctuations in power use, such as when a solar panel's output suddenly drops because a cloud passes overhead. The second part acts like a long-term memory, remembering the big picture of how a household's energy habits change over days and weeks. By stitching these two capabilities together, the system can see both the immediate details and the broader story of energy consumption.

To test if this new framework actually works, the team ran extensive simulations using real-world data from hundreds of homes in Australia, including records from the Ausgrid network and the Sydney region. They compared their new method against older, more traditional approaches that relied on a central server to collect and process all the information. The results showed that the traditional methods struggled significantly when faced with the messy, unpredictable reality of different households. They often failed to predict the highest peaks in energy demand, which are the most critical moments for grid stability. In contrast, the new decentralised system, with its dual-engine learning model, proved to be remarkably accurate. In the Sydney dataset, it achieved a prediction accuracy score of 0.97, meaning it captured almost all the variations in real energy use, and its average error rate was a tiny 0.74 percent.

This success suggests that the future of managing local power grids does not require a central brain that knows everything about everyone. Instead, the network can function as a collective of intelligent neighbors, each protecting their own privacy while contributing to a shared, highly reliable forecast. The researchers demonstrated that this method works even when the data is messy and the conditions are constantly changing, such as during sudden weather shifts or when solar generation varies wildly. While the system requires more time to train than simpler models, the trade-off is a level of precision and security that older methods cannot match. The study confirms that it is possible to build a smart, adaptive energy grid that respects user privacy and handles the complex, dynamic nature of modern renewable energy, paving the way for more resilient and efficient local power systems.

Technical Summary: An Adaptive Decentralised Federated Learning Framework for Load Forecasting in Dynamic Microgrid Environments

1. Problem Statement

Modern microgrids rely on Energy Management Systems (EMS) to balance intermittent renewable generation (e.g., solar) with dynamic, heterogeneous household demand. Effective load forecasting is critical for this balance; however, traditional approaches face three primary challenges:

Data Heterogeneity: Consumer load profiles vary significantly due to diverse behaviors, appliance usage, and renewable generation patterns. Centralized deep learning models often struggle to capture these long-term temporal dependencies across such varied profiles.
Privacy and Security: Centralizing raw consumption data from smart meters creates significant privacy vulnerabilities, exposing sensitive user behavioral patterns (e.g., occupancy schedules).
Systemic Limitations of Centralized Federated Learning (CFL): While CFL protects privacy by keeping data local, it relies on a central server for model aggregation. This introduces single points of failure, communication bottlenecks, and memory overload on resource-constrained edge hardware. Furthermore, standard LSTM models used in these settings often suffer from "forgetting" mechanisms that degrade long-term trend capture and struggle with instantaneous time-slice fluctuations.

2. Methodology

The authors propose a Decentralised Federated Learning (DFL) framework integrated with a hybrid TCN-BiLSTM architecture. This approach eliminates the central server dependency and enhances temporal feature extraction.

2.1 Hybrid Architecture (TCN-BiLSTM)

The core forecasting engine combines two neural network components to address different temporal scales:

Temporal Convolutional Networks (TCN): Utilizes dilated causal convolutions to capture instantaneous time-slice fluctuations and micro-patterns (e.g., 30-minute solar fluctuations or appliance surges). The TCN preserves temporal resolution while scanning fixed-length time windows.
Bidirectional LSTMs (BiLSTM): Processes the features extracted by the TCN using parallel forward and backward chains. This allows the model to retain macro temporal dependencies and understand trends in both directions (e.g., how morning conditions affect afternoon generation), mitigating the "forgetting" issues of standard LSTMs.

2.2 Decentralised Federated Learning (DFL) with Gossip Protocol

Instead of a central aggregator, the framework employs a peer-to-peer (P2P) gossip protocol:

Local Training: Each household (client) trains its local TCN-BiLSTM model on private data (solar generation and grid consumption).
Peer-to-Peer Aggregation: Clients periodically exchange model parameters ( $\theta$ ) with a subset of neighbors ( $N_i$ ) rather than a central server.
Update Rule: The update follows a gossip rule: $\theta_i \leftarrow \theta_i + \eta(\theta_j - \theta_i) + \epsilon$ , where $\eta$ is the mixing rate and $\epsilon$ is calibrated noise for differential privacy.
Convergence: The system aims to minimize local loss functions while bounding the gradient variance ( $\sigma^2$ ) and ensuring model consensus ( $\delta$ ) across the network without a central coordinator.

2.3 Mathematical Formalization

The framework formulates the load forecasting problem as a bounded optimization objective. It minimizes the global loss $L(\theta)$ subject to constraints on gradient variance and network connectivity. The system incorporates differential privacy mechanisms (Gaussian noise injection) to prevent the reconstruction of raw data from model updates.

3. Key Contributions

The paper outlines the following specific contributions:

Adaptive DFL Framework: Development of a decentralized framework that reduces reliance on centralized power networks, enabling collaboration among distributed energy resources while preserving data privacy.
Hybrid DFL-TCN-BiLSTM Model: Design of a novel hybrid model specifically for dynamic microgrids. It integrates TCN for short-term feature extraction and BiLSTM for long-term trend retention, improving the accuracy of short-term load forecasting to support supply-demand balancing.
Robustness to Heterogeneity: The framework is designed to handle heterogeneous time-series forecasting patterns, enabling efficient decision-making for EMS in dynamic environments where consumer behaviors and renewable generation fluctuate.

4. Experimental Results

The framework was validated using two Australian datasets: Ausgrid (300 households, 2012–2013) and NSW-Sydney (2019–2020). The proposed model was compared against baselines: Centralized FL with LSTM (CFL-LSTM) and Decentralized FL with LSTM (DFL-LSTM).

Performance Metrics: The DFL-TCN-BiLSTM framework achieved an $R^2$ of 0.97 and a MAPE of 0.74% on the NSW-Sydney dataset.
Comparison:
- vs. CFL-LSTM: The centralized baseline struggled with peak load surges, showing lower $R^2$ (0.78–0.82) and higher MAPE (up to 9.1%). It also suffered from single-point failure risks and communication bottlenecks.
- vs. DFL-LSTM: While DFL-LSTM improved over CFL by removing the central bottleneck, it still faced limitations in capturing multi-scale temporal features. The hybrid DFL-TCN-BiLSTM further reduced MAPE (to 3.2–3.8% in specific client tests) and RMSE, demonstrating superior tracking of extreme load peaks (>50 kWh).
Convergence: The study noted that while the hybrid model requires longer training times compared to simpler baselines (e.g., ~10 hours on TPU vs. ~4 hours for CFL-LSTM), it offers superior generalization and stability across heterogeneous client profiles.

5. Significance and Claims

The authors claim that this research provides a secure, scalable, and highly reliable forecasting engine for dynamic microgrid environments. The significance of the work lies in:

Privacy Preservation: By utilizing DFL, raw consumption data never leaves the local smart meter, addressing regulatory and privacy concerns inherent in centralized data collection.
Edge Suitability: The framework is designed for resource-constrained edge EMS hardware, eliminating the memory overload associated with central servers.
Handling Variability: The hybrid architecture effectively captures both instantaneous fluctuations (via TCN) and macro trends (via BiLSTM), addressing the limitations of standard LSTM models in capturing long-term dependencies and specific time-slice patterns.

The paper concludes that while the model introduces complexity and longer convergence times, the trade-off yields a robust system capable of managing the variability of renewable energy and heterogeneous consumer demands without compromising user privacy. Future work is identified as extending this simulation to live microgrid EMS deployments with client-specific personalization.