Periodic Reporting for period 1 - ANSWER (A heterogeneous distributed prediction model for wind-solar energy production)
Periodo di rendicontazione: 2024-02-01 al 2026-03-31
Sintesi del contesto e degli obiettivi generali del progetto
Wind and solar power are central to the European Green Deal and the EU's target of climate neutrality by 2050. However, the variable nature of renewable generation makes it difficult for grid operators to balance supply and demand, and extreme weather events such as blade icing can suddenly reduce wind-farm output and damage turbines. Better forecasting and monitoring tools are therefore essential to increase the share of renewables while keeping the power system reliable and affordable.
The ANSWER project (Grant Agreement No. 101111188) developed a new generation of machine-learning methods to analyse renewable-energy data that are distributed across many locations, devices, and owners. Rather than forcing every wind farm or solar plant to share its raw operational data, the project designed ways for local "client" models to learn from their own data and for a central server to combine only the learned patterns into a shared global model. This protects commercial and personal privacy, respects data-ownership boundaries, and still lets operators benefit from collective knowledge.
The overall objective was to build a heterogeneous distributed prediction framework that improves the accuracy and reliability of renewable-energy production forecasting. The project also targeted lightweight client models that can run on low-power edge devices, and server-side aggregation methods that remain accurate when client data differ in climate, geography, and sensor quality.
Work performed and main achievements
The project delivered six peer-reviewed scientific advances. For wind-power forecasting, LightWind provides a very small but accurate spatiotemporal model (184,000 parameters) that outperforms much larger methods on real wind-farm benchmarks while running fast enough for edge devices. For blade icing detection—a major safety and efficiency risk in cold climates—three complementary methods were developed: MW-DENet uses multiscale wavelet analysis and a dual-expert classifier to catch rare icing events even when iced samples are outnumbered 100 to 1; DGMEHIT combines local signal dynamics with global statistical features so that a model trained in one region works well in geographically different regions; and FedIce enables multiple wind farms to learn an icing detector together without sharing raw data, using only 20-50% of the usual labelled data. To handle the imperfect sensor data common in the field, the project also produced SPECTRA, an information-theoretic method for classifying incomplete and irregular time series, and an end-to-end joint imputation-and-classification model for missing sensor values.
Results and potential impacts
The new methods consistently exceed current state-of-the-art performance. LightWind reduces forecast errors on standard wind-farm datasets; DGMEHIT improves detection F1-score by 6.4% and Matthews correlation by 12.7%; FedIce improves federated icing detection by over 95% in key metrics with limited labels; and SPECTRA sets new records on public benchmarks for irregular time-series classification. Because the models are lightweight and privacy-preserving, they can be deployed directly on wind-farm controllers or edge computers, reducing data-transmission costs and enabling faster responses. The work supports EU energy and climate policy by making renewable generation more predictable and reliable, lowering balancing costs, improving turbine safety, and accelerating the clean-energy transition.
The ANSWER project (Grant Agreement No. 101111188) developed a new generation of machine-learning methods to analyse renewable-energy data that are distributed across many locations, devices, and owners. Rather than forcing every wind farm or solar plant to share its raw operational data, the project designed ways for local "client" models to learn from their own data and for a central server to combine only the learned patterns into a shared global model. This protects commercial and personal privacy, respects data-ownership boundaries, and still lets operators benefit from collective knowledge.
The overall objective was to build a heterogeneous distributed prediction framework that improves the accuracy and reliability of renewable-energy production forecasting. The project also targeted lightweight client models that can run on low-power edge devices, and server-side aggregation methods that remain accurate when client data differ in climate, geography, and sensor quality.
Work performed and main achievements
The project delivered six peer-reviewed scientific advances. For wind-power forecasting, LightWind provides a very small but accurate spatiotemporal model (184,000 parameters) that outperforms much larger methods on real wind-farm benchmarks while running fast enough for edge devices. For blade icing detection—a major safety and efficiency risk in cold climates—three complementary methods were developed: MW-DENet uses multiscale wavelet analysis and a dual-expert classifier to catch rare icing events even when iced samples are outnumbered 100 to 1; DGMEHIT combines local signal dynamics with global statistical features so that a model trained in one region works well in geographically different regions; and FedIce enables multiple wind farms to learn an icing detector together without sharing raw data, using only 20-50% of the usual labelled data. To handle the imperfect sensor data common in the field, the project also produced SPECTRA, an information-theoretic method for classifying incomplete and irregular time series, and an end-to-end joint imputation-and-classification model for missing sensor values.
Results and potential impacts
The new methods consistently exceed current state-of-the-art performance. LightWind reduces forecast errors on standard wind-farm datasets; DGMEHIT improves detection F1-score by 6.4% and Matthews correlation by 12.7%; FedIce improves federated icing detection by over 95% in key metrics with limited labels; and SPECTRA sets new records on public benchmarks for irregular time-series classification. Because the models are lightweight and privacy-preserving, they can be deployed directly on wind-farm controllers or edge computers, reducing data-transmission costs and enabling faster responses. The work supports EU energy and climate policy by making renewable generation more predictable and reliable, lowering balancing costs, improving turbine safety, and accelerating the clean-energy transition.
Lavoro eseguito dall’inizio del progetto fino alla fine del periodo coperto dalla relazione e principali risultati finora ottenuti
The ANSWER project (Grant Agreement No. 101111188) was a 24-month Marie Skłodowska-Curie Actions Postdoctoral Fellowship running from 1 February 2024 to 31 January 2026. Its scientific and technical work was organised around three research objectives and four research work packages: model specification, lightweight client forecasting models, heterogeneous server-side aggregation, and deployment/evaluation.
Objective 1 - Heterogeneous distributed forecasting framework. The project developed federated and distributed learning methods that allow multiple renewable-energy sites to collaborate without sharing raw operational data. The main achievement is FedIce, a prototype-enhanced federated semi-supervised learning framework for blade icing detection across distributed wind farms. FedIce mitigates data scarcity, severe class imbalance, and Non-IID data heterogeneity by learning class prototypes locally and aggregating only those prototypes at the server. On real-world data from 20 wind turbines across two wind farms, FedIce improved mFβ by 95.58% and mBA by 33.25% over the strongest baseline (FedMatch) when only 20% of data were labelled, while remaining competitive with fully supervised federated methods that use far more labels.
Objective 2 - Lightweight flexible client model. The project produced LightWind, a lightweight spatiotemporal framework for wind power forecasting. LightWind combines a Shuffled Group Temporal Convolutional Network (SG-TCN) with a Spatial Fusion-based Transformer (SF-Former) and a cosine-similarity adjacency matrix for turbine-to-turbine interactions. It supports both single-step and multistep forecasting across short-term (8 h -> 1 h), medium-term (24 h -> 6 h), and long-term (72 h -> 24 h) horizons without retraining for each horizon. On the SDWPF and Penmanshiel benchmarks, LightWind achieved the lowest MAE, RMSE, and SMAPE among seven state-of-the-art baselines, with statistical significance confirmed by Wilcoxon signed-rank tests. The model has only 184.36 K parameters, 53.56 M FLOPs, 12.31 MB peak memory, and 0.2431 s inference latency, making it suitable for resource-constrained edge devices. Complementary client-side robustness was provided by SPECTRA (AAAI-26), an information-theoretic framework for irregular time-series classification under missing data and class imbalance, and by an end-to-end incomplete time-series classification model (ESWA 2025) that jointly imputes missing values and classifies.
Objective 3 - Server-side aggregation with high generalisation. The project developed DGMEHIT, a domain-generalisable blade icing detection method that mixes Euclidean deep features with Hilbert-space statistics regularised by Maximum Mean Discrepancy (MMD). DGMEHIT improves cross-regional transferability: trained on Inner Mongolia data and tested on Shaanxi and Henan wind farms, it outperformed existing methods with F1 and MCC improvements of up to 11.6% and 20.8%, respectively. This demonstrates that server-level models can generalise across heterogeneous clients and climates without sharing raw data.
Additional technical achievements include MW-DENet (IEEE PSET 2025), a multiscale wavelet dual-expert network for imbalanced blade icing detection that achieved F1 scores of 95.45%, 87.21%, and 81.05% on datasets with imbalance ratios 1:20, 1:50, and 1:100.
The planned Bornholm living-lab demonstration could not be completed within the project period. Evaluation was instead carried out on public wind-farm benchmarks (SDWPF, Penmanshiel), proprietary industrial blade-icing datasets (GoldWind/Inner Mongolia, Shaanxi, Henan), standard time-series classification benchmarks (UCR, UEA, PAM, P12, P19), and an online monitoring experiment of DGMEHIT on 329.28 hours of continuous SCADA data from a wind turbine in Inner Mongolia, which achieved 100% detection accuracy during icing periods with almost no false alarms.
Objective 1 - Heterogeneous distributed forecasting framework. The project developed federated and distributed learning methods that allow multiple renewable-energy sites to collaborate without sharing raw operational data. The main achievement is FedIce, a prototype-enhanced federated semi-supervised learning framework for blade icing detection across distributed wind farms. FedIce mitigates data scarcity, severe class imbalance, and Non-IID data heterogeneity by learning class prototypes locally and aggregating only those prototypes at the server. On real-world data from 20 wind turbines across two wind farms, FedIce improved mFβ by 95.58% and mBA by 33.25% over the strongest baseline (FedMatch) when only 20% of data were labelled, while remaining competitive with fully supervised federated methods that use far more labels.
Objective 2 - Lightweight flexible client model. The project produced LightWind, a lightweight spatiotemporal framework for wind power forecasting. LightWind combines a Shuffled Group Temporal Convolutional Network (SG-TCN) with a Spatial Fusion-based Transformer (SF-Former) and a cosine-similarity adjacency matrix for turbine-to-turbine interactions. It supports both single-step and multistep forecasting across short-term (8 h -> 1 h), medium-term (24 h -> 6 h), and long-term (72 h -> 24 h) horizons without retraining for each horizon. On the SDWPF and Penmanshiel benchmarks, LightWind achieved the lowest MAE, RMSE, and SMAPE among seven state-of-the-art baselines, with statistical significance confirmed by Wilcoxon signed-rank tests. The model has only 184.36 K parameters, 53.56 M FLOPs, 12.31 MB peak memory, and 0.2431 s inference latency, making it suitable for resource-constrained edge devices. Complementary client-side robustness was provided by SPECTRA (AAAI-26), an information-theoretic framework for irregular time-series classification under missing data and class imbalance, and by an end-to-end incomplete time-series classification model (ESWA 2025) that jointly imputes missing values and classifies.
Objective 3 - Server-side aggregation with high generalisation. The project developed DGMEHIT, a domain-generalisable blade icing detection method that mixes Euclidean deep features with Hilbert-space statistics regularised by Maximum Mean Discrepancy (MMD). DGMEHIT improves cross-regional transferability: trained on Inner Mongolia data and tested on Shaanxi and Henan wind farms, it outperformed existing methods with F1 and MCC improvements of up to 11.6% and 20.8%, respectively. This demonstrates that server-level models can generalise across heterogeneous clients and climates without sharing raw data.
Additional technical achievements include MW-DENet (IEEE PSET 2025), a multiscale wavelet dual-expert network for imbalanced blade icing detection that achieved F1 scores of 95.45%, 87.21%, and 81.05% on datasets with imbalance ratios 1:20, 1:50, and 1:100.
The planned Bornholm living-lab demonstration could not be completed within the project period. Evaluation was instead carried out on public wind-farm benchmarks (SDWPF, Penmanshiel), proprietary industrial blade-icing datasets (GoldWind/Inner Mongolia, Shaanxi, Henan), standard time-series classification benchmarks (UCR, UEA, PAM, P12, P19), and an online monitoring experiment of DGMEHIT on 329.28 hours of continuous SCADA data from a wind turbine in Inner Mongolia, which achieved 100% detection accuracy during icing periods with almost no false alarms.
Progressi oltre lo stato dell’arte e potenziale impatto previsto (incluso l’impatto socioeconomico e le implicazioni sociali più ampie del progetto fino ad ora)
The ANSWER project produced six peer-reviewed scientific publications that advance the state of the art in distributed, robust, and efficient machine learning for renewable energy.
1. Information-theoretic irregular time-series classification (SPECTRA, AAAI-26). SPECTRA is the first framework to formally model and exploit the coupling between missingness patterns and class imbalance in irregular time series. It introduces Missingness-Aware Frequency Filtering (MAFF), a Missingness Pattern Encoder (MPE), and Category-Guided Feature Refinement (CGFR) with prototype-based learning on an information-geometric manifold. SPECTRA achieves state-of-the-art results on the P12, P19, and PAM benchmarks, outperforming 14 baselines including MTS-Former, ViTST, MuSiCNet, Raindrop, and WarpFormer.
2. Lightweight wind power forecasting (LightWind, IEEE Sensors Journal 2025). LightWind demonstrates that high-accuracy wind-farm forecasting does not require large models. Its grouped/shuffled temporal convolutions, lightweight spatial transformer, and end-stage temporal compression yield the best MAE, RMSE, and SMAPE on SDWPF and Penmanshiel across multiple horizons, while using only 184.36 K parameters and 53.56 M FLOPs. This is a significant advance over deep graph-neural-network and transformer alternatives that are too large for edge devices.
3. Domain-generalisable blade icing detection (DGMEHIT, IEEE Transactions on Industrial Informatics 2026). DGMEHIT combines local sequential features in Euclidean space with global statistical features in a reproducing kernel Hilbert space, regularised by MMD. It improves F1 by 6.4% and MCC by 12.7% over state-of-the-art blade icing methods, and shows strong cross-regional generalisation from Inner Mongolia to Shaanxi and Henan, addressing a practical barrier to deploying AI models across geographically distributed wind farms.
4. Federated semi-supervised blade icing detection (FedIce, IEEE Internet of Things Journal 2025). FedIce advances privacy-preserving distributed learning by using prototype aggregation to handle label scarcity, class imbalance, and Non-IID heterogeneity simultaneously. With only 20% labelled data, it improves mFβ by 95.58% over FedMatch; with 30% labels it reaches mFβ 84.99 and mBA 93.07 competitive with supervised federated learning that uses full labels.
5. Multiscale imbalanced blade icing detection (MW-DENet, IEEE PSET 2025). MW-DENet integrates 3-level discrete wavelet decomposition with a dual-expert classifier that includes an anomaly-focused learner trained exclusively on icing samples. It achieves F1 scores of 95.45%, 87.21%, and 81.05% on datasets with 1:20, 1:50, and 1:100 imbalance ratios, substantially reducing false-negative rates for this rare but critical event.
6. End-to-end incomplete time-series classification (ESWA 2025). The project developed a joint imputation-and-classification model for missing time-series data. It achieved an average accuracy of 90.3% on four real-world incomplete datasets and an 8.67% average improvement over the previous state of the art across 68 UCR datasets, with a very compact model (57.98 K parameters, 0.27 MB size).
Potential impacts and further uptake. The results provide ready-to-use algorithms for wind-farm forecasting, icing detection, and federated analytics. Key needs for further uptake include: (i) demonstration on additional wind-solar hybrid sites and living-lab environments; (ii) integration with industrial SCADA/IoT platforms and real-time edge hardware; (iii) patenting and commercialisation support through technology-transfer offices; (iv) standardisation of data interfaces for cross-farm federated learning; and (v) engagement with TSOs/DSOs and wind-farm operators to validate operational benefits and define acceptance criteria.
1. Information-theoretic irregular time-series classification (SPECTRA, AAAI-26). SPECTRA is the first framework to formally model and exploit the coupling between missingness patterns and class imbalance in irregular time series. It introduces Missingness-Aware Frequency Filtering (MAFF), a Missingness Pattern Encoder (MPE), and Category-Guided Feature Refinement (CGFR) with prototype-based learning on an information-geometric manifold. SPECTRA achieves state-of-the-art results on the P12, P19, and PAM benchmarks, outperforming 14 baselines including MTS-Former, ViTST, MuSiCNet, Raindrop, and WarpFormer.
2. Lightweight wind power forecasting (LightWind, IEEE Sensors Journal 2025). LightWind demonstrates that high-accuracy wind-farm forecasting does not require large models. Its grouped/shuffled temporal convolutions, lightweight spatial transformer, and end-stage temporal compression yield the best MAE, RMSE, and SMAPE on SDWPF and Penmanshiel across multiple horizons, while using only 184.36 K parameters and 53.56 M FLOPs. This is a significant advance over deep graph-neural-network and transformer alternatives that are too large for edge devices.
3. Domain-generalisable blade icing detection (DGMEHIT, IEEE Transactions on Industrial Informatics 2026). DGMEHIT combines local sequential features in Euclidean space with global statistical features in a reproducing kernel Hilbert space, regularised by MMD. It improves F1 by 6.4% and MCC by 12.7% over state-of-the-art blade icing methods, and shows strong cross-regional generalisation from Inner Mongolia to Shaanxi and Henan, addressing a practical barrier to deploying AI models across geographically distributed wind farms.
4. Federated semi-supervised blade icing detection (FedIce, IEEE Internet of Things Journal 2025). FedIce advances privacy-preserving distributed learning by using prototype aggregation to handle label scarcity, class imbalance, and Non-IID heterogeneity simultaneously. With only 20% labelled data, it improves mFβ by 95.58% over FedMatch; with 30% labels it reaches mFβ 84.99 and mBA 93.07 competitive with supervised federated learning that uses full labels.
5. Multiscale imbalanced blade icing detection (MW-DENet, IEEE PSET 2025). MW-DENet integrates 3-level discrete wavelet decomposition with a dual-expert classifier that includes an anomaly-focused learner trained exclusively on icing samples. It achieves F1 scores of 95.45%, 87.21%, and 81.05% on datasets with 1:20, 1:50, and 1:100 imbalance ratios, substantially reducing false-negative rates for this rare but critical event.
6. End-to-end incomplete time-series classification (ESWA 2025). The project developed a joint imputation-and-classification model for missing time-series data. It achieved an average accuracy of 90.3% on four real-world incomplete datasets and an 8.67% average improvement over the previous state of the art across 68 UCR datasets, with a very compact model (57.98 K parameters, 0.27 MB size).
Potential impacts and further uptake. The results provide ready-to-use algorithms for wind-farm forecasting, icing detection, and federated analytics. Key needs for further uptake include: (i) demonstration on additional wind-solar hybrid sites and living-lab environments; (ii) integration with industrial SCADA/IoT platforms and real-time edge hardware; (iii) patenting and commercialisation support through technology-transfer offices; (iv) standardisation of data interfaces for cross-farm federated learning; and (v) engagement with TSOs/DSOs and wind-farm operators to validate operational benefits and define acceptance criteria.