Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Task-Aware Machine Unlearning and Its Application in Load Forecasting

Wangkun Xu, Fei Teng

TL;DR

This paper tackles data privacy in load forecasting by enabling machine unlearning to remove the influence of selected training data from an already trained forecaster. It builds a framework based on influence functions and Newton updates to quantify and implement data removal, and introduces two enhancements—Performance-Aware MU (PAMU) and Task-Aware MU (TAMU)—to balance unlearning with downstream generator-dispatch costs. TAMU formulates a tri-level optimization and proves gradient existence to enable sample reweighting that aligns unlearning with power-system operation, while PAMU reweights remaining data to mitigate performance loss. Experiments on linear, CNN, and MLP-Mixer forecasters with a Texas-based dataset show that PAMU/TAMU can reduce operational costs while managing the completeness of unlearning, with code available for replication.

Abstract

Data privacy and security have become a non-negligible factor in load forecasting. Previous researches mainly focus on training stage enhancement. However, once the model is trained and deployed, it may need to `forget' (i.e., remove the impact of) part of training data if the these data are found to be malicious or as requested by the data owner. This paper introduces the concept of machine unlearning which is specifically designed to remove the influence of part of the dataset on an already trained forecaster. However, direct unlearning inevitably degrades the model generalization ability. To balance between unlearning completeness and model performance, a performance-aware algorithm is proposed by evaluating the sensitivity of local model parameter change using influence function and sample re-weighting. Furthermore, we observe that the statistical criterion such as mean squared error, cannot fully reflect the operation cost of the downstream tasks in power system. Therefore, a task-aware machine unlearning is proposed whose objective is a trilevel optimization with dispatch and redispatch problems considered. We theoretically prove the existence of the gradient of such an objective, which is key to re-weighting the remaining samples. We tested the unlearning algorithms on linear, CNN, and MLP-Mixer based load forecasters with a realistic load dataset. The simulation demonstrates the balance between unlearning completeness and operational cost. All codes can be found at https://github.com/xuwkk/task_aware_machine_unlearning.

Task-Aware Machine Unlearning and Its Application in Load Forecasting

TL;DR

This paper tackles data privacy in load forecasting by enabling machine unlearning to remove the influence of selected training data from an already trained forecaster. It builds a framework based on influence functions and Newton updates to quantify and implement data removal, and introduces two enhancements—Performance-Aware MU (PAMU) and Task-Aware MU (TAMU)—to balance unlearning with downstream generator-dispatch costs. TAMU formulates a tri-level optimization and proves gradient existence to enable sample reweighting that aligns unlearning with power-system operation, while PAMU reweights remaining data to mitigate performance loss. Experiments on linear, CNN, and MLP-Mixer forecasters with a Texas-based dataset show that PAMU/TAMU can reduce operational costs while managing the completeness of unlearning, with code available for replication.

Abstract

Data privacy and security have become a non-negligible factor in load forecasting. Previous researches mainly focus on training stage enhancement. However, once the model is trained and deployed, it may need to `forget' (i.e., remove the impact of) part of training data if the these data are found to be malicious or as requested by the data owner. This paper introduces the concept of machine unlearning which is specifically designed to remove the influence of part of the dataset on an already trained forecaster. However, direct unlearning inevitably degrades the model generalization ability. To balance between unlearning completeness and model performance, a performance-aware algorithm is proposed by evaluating the sensitivity of local model parameter change using influence function and sample re-weighting. Furthermore, we observe that the statistical criterion such as mean squared error, cannot fully reflect the operation cost of the downstream tasks in power system. Therefore, a task-aware machine unlearning is proposed whose objective is a trilevel optimization with dispatch and redispatch problems considered. We theoretically prove the existence of the gradient of such an objective, which is key to re-weighting the remaining samples. We tested the unlearning algorithms on linear, CNN, and MLP-Mixer based load forecasters with a realistic load dataset. The simulation demonstrates the balance between unlearning completeness and operational cost. All codes can be found at https://github.com/xuwkk/task_aware_machine_unlearning.
Paper Structure (36 sections, 3 theorems, 39 equations, 9 figures, 3 tables)

This paper contains 36 sections, 3 theorems, 39 equations, 9 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Data Privacy and Security in Load Forecasting
  3. Machine Unlearning
  4. Research Gaps
  5. Unlearning Completeness vs Model Performance
  6. Physical Meaning of Power System
  7. Contributions
  8. Machine Unlearning for Load Forecasting
  9. Parametric Load Forecasting Model
  10. Influence Function
  11. Machine Unlearning Algorithm
  12. Performance-aware Machine Unlearning
  13. Task-aware Machine Unlearning
  14. Formulation and Algorithm
  15. Extension to Neural Network based Load Forecaster
  16. ...and 21 more sections

Key Result

Proposition 1

The gradients $\partial \bm{p}_1^\star(\hat{\bm{y}}^i)\slash \partial \hat{\bm{y}}$ and $\partial \bm{p}_2^\star(\bm{P_g}^i) \slash \partial \bm{P}_g$ exist, which do not depend on $\hat{\bm{y}}^i$ and $\bm{P_g}^i$, respectively, if 1). $\bm{Q}$ is positive definite, $\bm{c}_{ls2}$ and $\bm{c}_{gs2}

Figures (9)

  • Figure 1: A workflow of machine unlearning. The data removal request can be made by privacy and security concerns. An unlearning algorithm is developed to update the forecaster with the role of power system operation being considered.
  • Figure 2: The structure of tri-level optimization \ref{['eq:obj_spo']} viewed as layers in the forward pass. The gradients used in \ref{['eq:cost_chain_rule']} are highlighted in red.
  • Figure 3: Structure of NN based load forecaster and feature extractor. All the layers except for the Flatten Layer and the Linear Layer 2 contain activations.
  • Figure 4: Performance of complete machine unlearning algorithm \ref{['eq:unlearning']} on remain (blue), unlearn (red) and test dataset (blue) of the linear load forecaster. The dotted curves report the performance of the original model and the solid curves are the performance of the unlearnt model.
  • Figure 5: Relationship on the influences of MSE, MAPE, and Cost criteria on the test dataset from the samples in remain dataset. The $r$ values are Pearson correlation coefficients.
  • ...and 4 more figures

Theorems & Definitions (4)

  • Proposition 1
  • Proposition 2
  • Proposition 3
  • proof