Federated edge intelligence for carbon emission forecasting using deep neural networks: a privacy-preserving approach

Accurate and timely carbon emission forecasting is a fundamental prerequisite for achieving carbon neutrality and optimizing energy management strategies in energy-intensive industrial parks. With the increasing deployment of Internet of Things (IoT) devices, massive amounts of energy data are generated at the network edge. However, traditional centralized cloud-based approaches for processing this data face critical challenges regarding data privacy, high transmission latency, and bandwidth consumption, particularly because industrial energy data often contains sensitive production information. To address these issues, this paper proposes a novel framework based on Federated Edge Intelligence (FEI) for collaborative carbon emission forecasting. Specifically, we design a distributed system where Deep Neural Networks (DNN) utilizing Long Short-Term Memory (LSTM) units are deployed on edge devices to capture the temporal dependencies of carbon emission data locally. A Federated Averaging (FedAvg) algorithm is employed to aggregate local model updates iteratively without sharing raw production data, thus breaking down data silos while preserving privacy. Furthermore, we formulate the global optimization problem and provide a detailed theoretical analysis of the LSTM gating mechanisms. Extensive experimental results on real-world energy consumption datasets demonstrate that the proposed method achieves a Root Mean Square Error (RMSE) of 0.096, which is comparable to centralized training, while reducing communication overhead by approximately 99% compared to raw data transmission. The results also validate the robustness of the proposed framework under non-IID data distributions typical in industrial environments.