1 of 21

Federated Continual Learning for Energy Forecasting in Residential Smart Metering Systems

CDS Research Group, 2nd year PhD student

ivonne.nunez@unibe.ch

Ivonne Núñez

​

2 of 21

Agenda

3rd Workshop on Intelligent IoT Services and Applications 2026

​

  1. Background-Challenges
  2. Previous Work
  3. Temporal-Similarity Based Client Clustering
  4. Forecasting Backbone
  5. Communication-Efficient Cluster-Level Aggregation
  6. Adaptive Personalization under Behavioral Drift
  7. Stability and Reliability Control
  8. Experimental Settings and Results
  9. Current work
  10. Next Steps

​

2

2

3 of 21

1. Background

Challenges

3rd Workshop on Intelligent IoT Services and Applications 2026

​

  • Heterogeneity: highly diverse consumption patterns degrade the performance of global models.
  • Temporal dynamics: changes due to seasonality, PV/EV, and habits require continuous adaptation without forgetting prior knowledge.
  • Efficiency: communication budgets and clients with limited connectivity require efficient and robust aggregation.
  • Reliability: not all local updates are stable; the system must be resistant to noise and variability.
  • Interpretability: predictions must be explained to build trust and support decision-making.

3

3

4 of 21

2. Previous Work -- PeFedTL

3rd Workshop on Intelligent IoT Services and Applications 2026

4

4

5 of 21

3. Previous Work -- ExFedL

3rd Workshop on Intelligent IoT Services and Applications 2026

5

5

6 of 21

4. Temporal-Similarity Based Client Clustering

Determination of the Distance Space:

Quality Evaluation:

​

3rd Workshop on Intelligent IoT Services and Applications 2026

Local Density Evaluation & Noise:

6

6

7 of 21

5. Forecasting Backbone

3rd Workshop on Intelligent IoT Services and Applications 2026

SHapley Additive exPlanations (SHAP):

 

Convolutional output of the TCN:

Attention weights for GAM:

Local-level explanations

Gobal-level explanations

Cluster consistency index

Normalized attribution profile

7

7

8 of 21

6. Communication-Efficient Cluster-Level Aggregation

3rd Workshop on Intelligent IoT Services and Applications 2026

The aggregation weight is defined as:

​

The confidence score is defined as:

 

 

 

 

8

8

9 of 21

7. Adaptive Personalization under Behavioral Drift

3rd Workshop on Intelligent IoT Services and Applications 2026

Drift detection with dynamic reassignment:

​

Lightweight personalization via FTL:

9

9

10 of 21

8. Stability and Reliability Control

3rd Workshop on Intelligent IoT Services and Applications 2026

Partition stability:

​

Explanation Stability:

We smooth both stability signals:

10

10

11 of 21

9. Experimental Settings and Results

3rd Workshop on Intelligent IoT Services and Applications 2026

Dataset

Data Type

Number of households and Time Period

Household Data (Germany)

Residential energy consumption, solar generation (60 minutes)

Six (6) residences. Starting point and quality vary, with gaps ranging from a few minutes to entire days.

CKW Smart Meter Data (Switzerland)

Residential energy consumption, aggregation at the municipal level, renewable energy (15 minutes)

Monthly data is continuously updated. 4959 smart meters have consumption and reported values over the full duration of 2 years.

Household energy consumption enriched with weather data in northeast of Mexico

Residential energy consumption and weather data (1 minute)

Energy consumption in a residence in the northeast region of Mexico over 14 months.

Datasets and Preprocessing

  • Clustering effectiveness
  • Forecasting accuracy
  • Efficiency and Robustness
  • Interpretability
  • Personalization and adaptability
  • Temporal reliability

Performance metrics

11

11

12 of 21

9. Experimental Settings and Results

3rd Workshop on Intelligent IoT Services and Applications 2026

Overall Forecasting Accuracy:

12

12

13 of 21

9. Experimental Settings and Results

3rd Workshop on Intelligent IoT Services and Applications 2026

Clustering Performance:

Density and Distribution of Clustering

MAE Distribution Across Clusters with and Without Clustering

13

13

14 of 21

9. Experimental Settings and Results

3rd Workshop on Intelligent IoT Services and Applications 2026

Efficiency and Robustness:

14

14

15 of 21

9. Experimental Settings and Results

3rd Workshop on Intelligent IoT Services and Applications 2026

Efficiency and Robustness:

15

15

16 of 21

9. Experimental Settings and Results

3rd Workshop on Intelligent IoT Services and Applications 2026

Personalization and Adaptability:

MAE Distribution for Cold-Start Clients

Behavioral Drift Detection And Reassignment Trigger

16

16

17 of 21

9. Experimental Settings and Results

3rd Workshop on Intelligent IoT Services and Applications 2026

Interpretability via SHAP:

Cluster-Level Contribution Share in Global SHAP Feature Importance

SHAP Local-Level Interpretability

17

17

18 of 21

9. Experimental Settings and Results

3rd Workshop on Intelligent IoT Services and Applications 2026

Temporal Reliability:

18

18

19 of 21

10. Current work

3rd Workshop on Intelligent IoT Services and Applications 2026

  • Dynamic clustering based on temporal similarity

​

  • Federated Continual Learning with stability control

​

  • Efficient and asynchrony-tolerant federated learning

​

  • Continuous personalization under drift

​

  • Temporally consistent explainability

Proposed contributions

19

19

20 of 21

11. Next Steps…

3rd Workshop on Intelligent IoT Services and Applications 2026

Formalize a Federated Continual Learning framework for smart metering by incorporating explicit continual learning. This includes memory and anti-forgetting mechanisms, dynamic clustering based on temporal behavior, multi-level control of stability and reliability, and native integration of explainability.

20

20

21 of 21

Thank you for your attention!���Questions?

CDS Research Group, 2nd year PhD student

ivonne.nunez@unibe.ch

Ivonne Núñez

​