Full Text Available
Note: Clicking the button above will open the full text document at the original institutional repository in a new window.
Historical Consistent Neural Networks for Wind Power Prediction
Thesis (PhD)--Stellenbosch University, 2026.
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Language: | English |
| Published: |
Stellenbosch : Stellenbosch University
2026
|
| Tags: |
No Tags, Be the first to tag this record!
|
| _version_ | 1867613944189812736 |
|---|---|
| access_status_str | Open Access |
| author | Rockefeller, Rockefeller |
| author2 | Bah, Bubacarr |
| author_browse | Bah, Bubacarr Rockefeller, Rockefeller |
| author_facet | Bah, Bubacarr Rockefeller, Rockefeller |
| author_sort | Rockefeller, Rockefeller |
| collection | Thesis |
| dc_rights_str_mv | Stellenbosch University |
| description | Thesis (PhD)--Stellenbosch University, 2026. |
| format | Thesis |
| id | oai:scholar.sun.ac.za:10019.1/135894 |
| institution | Stellenbosch University (South Africa) |
| language | English |
| last_indexed | 2026-06-10T12:44:10.803Z |
| license_str | Other — see source repository |
| provenance_str_mv | Harvested via OAI-PMH from SUNScholar — Stellenbosch University Repository |
| publishDate | 2026 |
| publishDateRange | 2026 |
| publishDateSort | 2026 |
| publisher | Stellenbosch : Stellenbosch University |
| publisherStr | Stellenbosch : Stellenbosch University |
| record_format | dspace |
| source_str | SUNScholar — Stellenbosch University Repository |
| spelling | oai:scholar.sun.ac.za:10019.1/135894 Historical Consistent Neural Networks for Wind Power Prediction Rockefeller, Rockefeller Bah, Bubacarr Marivate, Vukosi Zimmermann, Hans-Georg Stellenbosch University. Faculty of Science. Dept. of Applied Mathematics. Thesis (PhD)--Stellenbosch University, 2026. Rockefeller, R. 2026. Historical Consistent Neural Networks for Wind Power Prediction. Unpublished doctoral dissertation. Stellenbosch: Stellenbosch University [online]. Available: https://scholar.sun.ac.za/items/85fd7f2d-d033-452b-8c20-342006f8ca88 Modeling chaotic dynamical systems remains a fundamental challenge due to their inherent non-linearity, sensitivity to initial conditions, and long-range temporal dependencies. While Recurrent Neural Networks (RNNs) are widely used for such tasks, they often suffer from temporal inconsistency: relying on external inputs during training that are unavailable during forecasting. Historical Consistent Neural Networks (HCNNs) offer a principled alternative by embedding observed and unobserved variables into a unified state, ensuring the dynamical rules evolve consistently across past and future. While originally explored mainly in financial applications, this thesis advances HCNNs as scalable, physics-compatible architectures to model causal, high-dimensional, and nonlinear physical structures, with a primary emphasis on supporting reliable wind energy forecasts. The contributions of this work are threefold. First, a modular computational environment was designed to support rigorous HCNN experimentation, to ensure reproducibility across diverse architectural configurations. Second, novel HCNN variants were developed to address specific modeling challenges: (i) HCNN with partial Teacher Forcing (HCNN-pTF), to reduce dependence on observables, thereby forcing the model to align its internal representations with the autonomous behaviour of the physical structure; (ii) HCNN with an LSTM formulation (HCNN-LForm), incorporating an LSTM-style gating mechanism to balance memory retention with temporal dynamics; (iii) HCNN with Large Sparse transition matrix (HCNN-LSpa) which projects the system’s state into a significantly higher-dimensional space regularized by a structured sparsity approach. Third, these models were rigorously validated through a staged experimental protocol, moving from trajectory reconstruction in fully observable chaotic systems (Logistic map, Rabinovich-Fabrikant, Rössler, Lorenz) to forecasting in noisy, partially observable real-world data. In the latter regime, the framework’s ability to generate reliable wind energy forecasts demonstrates its practical utility, directly supporting: (i) turbine operation and control (e.g., blade/pitch decisions); (ii) short-term grid stability and dispatch; and (iii) longterm generation planning and investment decisions. In fully observable settings where the complete data history is available, fully unfolded HCNN architectures were used. Teacher forcing supports the learning and it yields to one learned starting point of the model from which reconstructed trajectories can be generated, hence forecasts. HCNN-pTF combined with an optimal adaptive dropout rate of 25% emerged as the superior architecture. While increasing noise levels significantly hindered reconstruction fidelity across most models, HCNN-pTF maintained dominant performance and stability, suggesting a better handling of data uncertainty when comparing to other variants, traditional RNN and LSTM baselines. Furthermore, ensemble training on canonical systems revealed a significant minimization of model uncertainty: distinct ensemble members consistently converged to the same physical solution regardless of random initialization. However, technical applications are characterized by long, high- dimensional data streams and cannot sustain “infinite” data histories due to fixed computational budgets, necessitating the use of limited sliding window. We introduce Abridged HCNNs which works by deploying the fixed past length architecture on a truncated part of the series and repositioning it during training. Here, the role of teacher forcing evolved into aligning the information flow in the model to every new repositioning, which is critical as there is no chance to train the network down to zero error. Building on these findings, the HCNN framework was applied to the practical challenge of wind energy forecasting in a specific location in South Africa. Wind modeling is a domain characterized by high-dimensional, noisy, and partially observed atmospheric dynamics. Leveraging innovative preprocessing strategies that embed diurnal cycles and physical constraints, HCNN-LSpa delivered stable, high-fidelity forecasts across short (12-hour) and day-ahead (24-hour) horizons, outperforming standard RNN and LSTM baselines. We attribute this superior generalization performance to three synergistic mechanisms. First, by projecting dynamics into a high-dimensional state space, the architecture effectively captured complex non-linear relationships within short temporal windows. Second, the standard Teacher Forcing (TF) worked to HCNN-LSpa’s advantage: it provided constant, high-fidelity ground truth corrections that allow the high-capacity network to focus entirely on learning the precise local transitions without the destabilizing effects of long-term drift. Third, the sparse connectivity acted as a structural regularizer, preventing the model from overfitting to specific window frames and ensuring that learned dynamics generalize across disjoint temporal slices. The use of observable-aware sparsity, which consists on a selective weight-retention mechanism, rather than by randomly assigning sparse locations, ensured the network retains direct sensitivity to contributions from true observables, consistently preserving the corrective feedback provided by the TF mechanism across the training period. These results establish HCNNs not merely as predictive tools or mathematical structures, but as interpretable, physics-compatible architectures. This stands in sharp contrast to traditional approaches: linear models lack the capacity to capture such complex dynamics. LSTM is effective at mitigating the long memory problem in learning, but at the cost of physical interpretability (Markov property). Our HCNN approach offers a causal understanding of the wind system, in contrast to a purely phenomenological data analysis, which would not allow the reconstruction of unobserved variables. Doctoral 2026-04-14T11:58:30Z 2026-04-14T11:58:30Z 2026-03 Thesis https://scholar.sun.ac.za/handle/10019.1/135894 en Stellenbosch University 319 pages application/pdf Stellenbosch : Stellenbosch University |
| spellingShingle | Rockefeller, Rockefeller Historical Consistent Neural Networks for Wind Power Prediction |
| title | Historical Consistent Neural Networks for Wind Power Prediction |
| title_full | Historical Consistent Neural Networks for Wind Power Prediction |
| title_fullStr | Historical Consistent Neural Networks for Wind Power Prediction |
| title_full_unstemmed | Historical Consistent Neural Networks for Wind Power Prediction |
| title_short | Historical Consistent Neural Networks for Wind Power Prediction |
| title_sort | historical consistent neural networks for wind power prediction |
| url | https://scholar.sun.ac.za/handle/10019.1/135894 |
| work_keys_str_mv | AT rockefellerrockefeller historicalconsistentneuralnetworksforwindpowerprediction |