Personalized Federated Learning for Noisy Multi-Edge Load Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing load prediction methods struggle with high variability and noise in edge computing environments, insufficient historical data, and limited generalization ability, leading to inaccurate and inefficient resource provisioning in edge systems.
Innovation Solution
A Multi-edge Cooperative Universal Framework for Load Prediction with Personalized Federated Deep Learning (MC-2PF) that utilizes site IDs to divide cooperative training regions, incorporates Savitzky-Golay filtering for noise reduction, and employs novel model control parameters to correct update directions, enabling adaptive and accurate load prediction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If classic neural networks (MLP, RBF) are used for load prediction, then the network structure is simple, but prediction accuracy deteriorates when facing highly-variable edge loads
Solution Approach 1:
The patent transitions from static neural network structures to dynamic Recurrent Neural Networks (RNNs) that can adapt to temporal patterns in load data. The RNN architecture allows the model to dynamically adjust its internal state based on sequential load observations, enabling it to capture time-dependent variations and achieve high prediction accuracy for highly-variable edge loads while maintaining reasonable structural complexity.
2Measurement precision
If Recurrent Neural Networks (RNNs) are used to capture time-series patterns, then prediction accuracy improves, but gradient vanishing or explosion occurs making it difficult to learn long-term memory dependencies
Solution Approach 1:
The patent employs Long Short-Term Memory (LSTM) networks, which introduce dynamic gating mechanisms (input gate, forget gate, output gate) that regulate information flow. These gates dynamically adjust the network's memory capacity based on the importance of temporal patterns, allowing the model to learn long-term dependencies without suffering from gradient vanishing or explosion, thus achieving both high prediction accuracy and learning stability.
3Device complexity
If deep learning models are trained independently by single edge servers, then model training is simple, but insufficient historical load data leads to poor prediction performance
Solution Approach 1:
The patent implements Federated Learning that merges the training capabilities of multiple edge servers into a unified collaborative system. Instead of each server training independently with limited local data, the system aggregates models and gradients across multiple servers, effectively combining their historical load data to train robust deep learning models. This collaboration enables each server to benefit from the collective data of the entire edge network, significantly improving prediction accuracy despite individual servers having insufficient historical data.
4Measurement precision
If Federated Learning is applied to combine data from multiple edge servers, then prediction accuracy improves, but client drift issue arises due to distinct data distribution
Solution Approach 1:
The patent introduces personalized learning mechanisms that allow each edge server to maintain its own local model parameters while participating in the federated learning process. This enables the system to capture both global patterns (through model aggregation) and local characteristics (through personalized adjustments), effectively addressing client drift caused by distinct data distributions. Each server's model is optimized for its specific local conditions while still benefiting from the collective knowledge of the network, improving both prediction accuracy and model generalization.
Data Source
AI summary
A Multi-edge Cooperative Universal Framework for Load Prediction with Personalized Federated Deep Learning adopts an FDL-based cooperative training manner in multi-edge environments, and uses site IDs of edge servers as the basis for dividing the regions of cooperative training. The edge servers within the same site are regarded as clients in FDL, and they conduct cooperative training with the cloud server (i.e., the parameter server in FDL), customize personalized models for each edge by independent control parameters and theoretically analyze the model convergence improvement.