本論文以資訊系統（IS）設計科學方法，探討自適應學習率元控制（meta-control）問題。其核心科學主張在於：當學習率的強化學習式動態元控制與網路結構適應在架構上明確分離時，相較於固定排程式啟發式演算法，在平穩、實證觀測的非平穩，以及前瞻性合成高再生能源預測情境中，能提供一種具情境依存性、具統計可再現性且治理導向的模型管理架構。此主張對應於本研究的架構假設：即當資料分佈之非平穩性增加時，自適應學習率元控制的相對價值將愈發顯著。其他貢獻（模型緊湊性、可解釋性、MLOps（機器學習維運）整合、邊緣部署可行性，以及相對於大型深度學習基線的競爭性表現），則被視為支持性設計知識啟示，用以建構和脈絡化此主要科學主張，而非取代之。
波動性再生能源的高滲透率改變了電力系統短期負載預測的運轉與決策意涵。隨著再生能源占比提高，系統運轉者日益需要預測淨負載（定義為系統總需求扣除波動性再生能源發電量後的剩餘需求），而不再僅聚焦於總負載預測。淨負載預測具有高度挑戰性，因其同時受到人類用電行為與天氣驅動之再生能源波動性影響；在高再生能源滲透率之系統中，常伴隨益發劇烈的升降載、更頻繁的低剩餘需求時段，以及更高的運轉壓力。這些特性不僅構成工程問題，也形成資訊管理問題：預測系統除須降低預測誤差外，亦應在實際運轉決策支援環境中，支援可重複的模型治理、具運算效率的再訓練流程，以及具可解釋性的模型管理決策。
本論文提出 CSI-SLFN-DDPG 元學習框架，用於電力系統淨負載預測。該框架結合緊湊型單一隱藏層前饋神經網路（single-hidden-layer feedforward network, SLFN）、CSI 啟發式結構自適應機制，以及基於深度確定性策略梯度（deep deterministic policy gradient, DDPG）的學習率元控制機制。此系統被設計為一項支援人工智慧決策支援之設計科學研究（design science research, DSR）工件。
本研究採用 2016 至 2020 年愛爾蘭電力系統淨負載資料進行實證評估，資料特徵包括落後期需求、風力發電、日曆特徵及衍生淨負載特徵。研究以 10 個隨機種子進行探索性實驗，並以 30 個隨機種子進行確認性配對實驗；模型穩健性則於三類情境下評估：歷史常規測試情境、實證觀測的歷史高壓力子集（265 個高再生能源時段，風力占比 ≥ 0.7），以及合成 2030 型高再生能源壓力情境，並採有界轉移觀點加以詮釋。合成 2030 型情境被定位為前瞻性的合成穩健性探測，它並非直接的運轉實證，因其資料分布部分建構於再生能源滲透率的前瞻性假設，而非完全來自歷史觀測的系統行為。因此，本研究將實證觀測的歷史高壓力情境定位為主要推論主張在實際部署運轉脈絡下的解釋錨點。
本研究以均方根誤差（root mean square error, RMSE）作為配對統計推論的主要評估指標；R^2 則僅作為補充性描述統計，因其為有界且非線性的指標，於高壓力時段可能呈現偏態分布。針對 RMSE 結果，本研究採用三種互補的配對推論程序進行多方法交叉檢視：配對 t 檢定、Wilcoxon 符號秩檢定，以及雙側符號翻轉置換檢定；並同時報告配對設計效果量，包括經 Hedges 小樣本修正的配對 Cohen's d_z、配對秩雙列相關，以及配對符號差值。另以 20,000 次重複抽樣估計百分位數 bootstrap 95% 信賴區間。
在 30 個隨機種子的確認性配對分析中，CSI-SLFN-DDPG 在三個評估情境的主要 RMSE 指標上，均相較於結構匹配的 CSI-SLFN-HEURISTIC 基線呈現方向一致、經多種配對推論程序支持的相對優勢。配對效果量依情境呈遞增模式，Cohen's d_z: 分別為 +0.464、+0.499 與 +0.540；其中，分布轉換最強的合成 2030 型完整測試情境具有最大的配對效果量（詳細的配對統計結果見第 5.4 節）。此結果與本研究的架構假設一致：在所設定的資料、模型與評估協定下，隨資料分布非平穩性提高，基於強化學習的自適應學習率元控制可能呈現較高的相對價值。
本論文據此提出雙層次的設計科學洞察。第一，在主要推論層次上，研究結果提供條件式證據，支持強化學習式學習率元控制在不同非平穩程度下相對效益可能遞增的架構假設。第二，在支持性設計知識層次上，CSI-SLFN-DDPG 作為一項緊湊、治理導向且具部署意識的預測架構，僅約含 300 個參數，平均每個隨機種子的訓練時間約為 18 秒，並可提供可稽核的學習率調整軌跡與模型容量演進紀錄。在三個評估情境的描述性比較中，其平均 RMSE 均低於參數規模顯著較大的深度學習基線 Transformer FFN（3,297 個參數）與 LSTM 型 FFN（6,433 個參數）。因此，該架構在預測精度、模型緊湊性、治理可稽核性與再訓練負擔之間，展現可供後續部署驗證與 MLOps 整合評估的實務潛力；實際系統調度者工作流程中的線上部署與實證驗證，則留待後續研究。

This dissertation studies adaptive learning-rate meta-control as an Information Systems (IS) design-science problem. Its central scientific claim is that reinforcement-learning-based meta-control of optimization dynamics—when architecturally separated from structural adaptation—delivers a regime-conditional, statistically reproducible, and governance-oriented architecture over fixed-schedule heuristics across stationary, authentic non-stationary, and forward-looking high-renewable forecasting regimes, consistent with the architectural hypothesis that adaptive learning-rate control becomes increasingly valuable as distributional non-stationarity increases. All other contributions—compactness, interpretability, MLOps integration, edge deployability, and competitive performance against larger deep-learning baselines—are treated as supporting design-knowledge implications that scaffold and contextualize this primary claim rather than replace it.
High penetration of variable renewable energy changes the operational meaning of short-term load forecasting. System operators increasingly require forecasts of net load—defined as demand minus variable renewable generation—rather than gross demand alone. Net load is difficult to forecast because it combines human consumption patterns with weather-driven renewable production, and because high-renewable systems exhibit sharper ramps, lower residual-demand periods, and more frequent operational stress. These characteristics create an information-management problem as much as an engineering problem: forecasting systems must not only minimize error but also support repeatable model governance, computationally efficient retraining, and interpretable model-management decisions in operational decision-support environments.
This dissertation proposes CSI-SLFN-DDPG, a meta-learning framework that combines a compact Single-Hidden-Layer Feedforward Network (SLFN) with CSI-inspired structural adaptation and DDPG-based learning-rate control for net-load forecasting. The system is designed as a Design Science Research (DSR) artifact for AI-enabled decision support. The empirical evaluation uses an Irish power-system net-load dataset (2016–2020) with lagged demand, wind generation, calendar, and derived net-load features. Robustness is assessed with both 10-seed exploratory and 30-seed confirmatory paired experimental designs across three regimes: Historical Normal Test conditions, Authentic Historical Stress subsets (265 empirically observed high-renewable hours with wind share ≥ 0.7), and Synthetic 2030-like high-renewable stress scenarios. The Synthetic 2030-like regime is interpreted as a forward-looking synthetic robustness probe rather than as direct operational evidence, because its distribution is partially constructed from projected renewable-penetration assumptions rather than observed historical system behavior; the Authentic Historical Stress regime is therefore designated the deployed-operation anchor for primary inferential claims.
Paired statistical inference is reported with RMSE designated as the primary inferential endpoint, while R² is provided as supplementary descriptive summary only because R² is bounded, nonlinear, and can be skewed on stress windows. Three complementary inferential procedures are triangulated on RMSE—paired t-test, Wilcoxon signed-rank test, and a two-sided sign-flip permutation test—together with paired-design effect sizes (paired Cohen's d_z with Hedges correction, matched-pairs rank-biserial correlation, and paired sign-delta) and percentile bootstrap 95% confidence intervals (20,000 resamples).
The 30-seed paired statistical analysis establishes a convergent and triangulated effectiveness profile for CSI-SLFN-DDPG across all three regimes on the primary RMSE endpoint, with the largest paired effect size observed under the regime exhibiting the strongest distributional transformation (Synthetic 2030-like Full Test, paired Cohen's d_z = +0.540). The Authentic Historical Stress regime serves as the deployed-operation anchor, while the Synthetic 2030-like regime is interpreted under bounded-transfer framing because its distribution is partially constructed from forward-projected renewable-penetration assumptions rather than observed system behavior. The proposed artifact is positioned as a governance-oriented forecasting architecture that exposes auditable learning-rate trajectories and capacity-evolution records, designed to support operational governance within MLOps workflows; empirical validation under live operator workflows is identified as future work.
Deployment-oriented recommendations are therefore anchored on the Authentic Historical Stress regime, where CSI-SLFN-DDPG attains a statistically convergent and practically meaningful paired advantage over the structurally matched heuristic baseline under empirically observed high-renewable conditions (full triangulated paired statistics in Section 5.4). Across all three evaluated regimes, CSI-SLFN-DDPG remains operationally attractive relative to substantially larger sequence-modelling baselines (Transformer FFN, 3,297 parameters; LSTM-style FFN, 6,433 parameters), while preserving a compact footprint of approximately 300 parameters, explicit capacity auditability, and low retraining overhead.
The dissertation therefore contributes a two-level scientific insight. First (primary inferential level), at the 30-seed confirmatory tier of the dual-tier paired protocol, CSI-SLFN-DDPG attains a convergent, triangulated paired advantage over the structurally matched CSI-SLFN-HEURISTIC baseline on the primary RMSE endpoint under each of the three evaluated regimes with paired effect sizes exhibiting a regime-ordered strengthening pattern (Cohen's d_z: +0.464 → +0.499→ +0.540), consistent with the architectural hypothesis that RL-based adaptive learning-rate control becomes increasingly valuable as distributional non-stationarity increases. Second (supporting design-knowledge level), the broader practical value of the artifact as a compact, governance-oriented, and deployment-aware forecasting architecture remains evident across all three regimes, with CSI-SLFN-DDPG (~300 parameters) attaining substantially lower mean RMSE than Transformer FFN (3,297 params) and LSTM-style FFN (6,433 params) under each regime .

These findings advance a design-science contribution with practical implications for AI-enabled decision support in high-renewable power systems. First, they provide an IS-oriented framing of adaptive forecasting as a governance-oriented artifact—exposing auditable learning-rate trajectories and capacity-evolution records—rather than as a prediction tool whose internal optimization dynamics are not directly inspectable. Second, they formalize a tractable MDP for learning-rate meta-control in a CSI-adaptive network that balances forecasting accuracy (mean RMSE 21.22 MW under Normal Test, 15.94 MW under Authentic Historical Stress) with architectural parsimony (~300 parameters, ~18 s mean per-seed training time). Third, they identify a regime-ordered strengthening pattern of RL-based adaptive meta-control under increasing distributional non-stationarity, while also showing that compact meta-learned architectures remain practically competitive against substantially larger sequence-modelling baselines under both empirically observed high-renewable conditions (Authentic Historical Stress) and forward-projected synthetic high-renewable conditions (Synthetic 2030-like Full Test, interpreted under bounded-transfer framing).

[1] IPCC. (2023). Climate change 2023: Synthesis report. Contribution of Working Groups I, II and III to the Sixth Assessment Report of the Intergovernmental Panel on Climate Change [Core Writing Team, H. Lee & J. Romero (Eds.)]. IPCC. https://doi.org/10.59327/IPCC/AR6-9789291691647.
[2] California Independent System Operator. (2013). What the duck curve tells us about managing a green grid. https://www.caiso.com/Documents/FlexibleResourcesHelpRenewables_FastFacts.pdf.
[3] Yang, D., Wang, W., Gueymard, C., Hong, T., Kleissl, J., Huang, J., Perez, M., Perez, R., Bright, J., Xia, X., van der Meer, D., & Peters, I. M. (2022). A review of solar forecasting, its dependence on atmospheric sciences and implications for grid integration: Towards carbon neutrality. Renewable and Sustainable Energy Reviews, 161, 112348. https://doi.org/10.1016/j.rser.2022.112348.
[4] International Energy Agency. (2018). Status of power system transformation 2018: Advanced power plant flexibility. https://iea.blob.core.windows.net/assets/ede9f1f7-282e-4a9b-bc97-a8f07948b63c/Status_of_Power_System_Transformation_2018.pdf. Retrived at: March 27, 2026
[5] Hevner, A.R., March, S.T., Park, J., & Ram, S. (2004). Design science in information systems research. MIS Quarterly, 28(1), 75–106. https://doi.org/10.2307/25148625.
[6] Howard, R.A. (1960). Dynamic Programming and Markov Processes. MIT Press, Cambridge, Massachusetts.
[7] Lillicrap, T.P., Hunt, J.J., Pritzel, A., Heess, N., Erez, T., Tassa, Y., Silver, D., & Wierstra, D. (2015). Continuous control with deep reinforcement learning. arXiv preprint arXiv:1509.02971. https://arxiv.org/abs/1509.02971.
[8] Jaderberg, M., Dalibard, V., Osindero, S., Czarnecki, W.M., Donahue, J., Razavi, A., Vinyals, O., Green, T., Dunning, I., Simonyan, K., Fernando, C., & Kavukcuoglu, K. (2017). Population Based Training of Neural Networks. ArXiv, abs/1711.09846.
[9] Kreuzberger, D., Kühl, N., & Hirschl, S. (2022). Machine learning operations (MLOps): Overview, definition, and architecture. arXiv preprint arXiv:2205.02302. https://doi.org/10.48550/arXiv.2205.02302.
[10] Hamad, Z., & Abdulrahman, I. (2022). Deep learning-based load forecasting considering data reshaping using MATLAB\Simulink. International Journal of Energy and Environmental Engineering, 13(2), 853–869. https://doi.org/10.1007/s40095-022-00480-x.
[11] Vivas, E., Allende-Cid, H., & Salas, R. (2020). A Systematic Review of Statistical and Machine Learning Methods for Electrical Power Forecasting with Reported MAPE Score. Entropy (Basel, Switzerland), 22(12), 1412. https://doi.org/10.3390/e22121412.
[12] Chan, J. W., & Yeo, C. K. (2024). A Transformer based approach to electricity load forecasting. The Electricity Journal, 37(2), 107370. https://doi.org/10.1016/j.tej.2024.107370.
[13] Chen, Y. Wu, Y., & An, Z. (2024). A novel hybrid model Informer-LSTM based on beluga whale optimization algorithm for stock price prediction. In Proceedings of the 2023 4th International Conference on Machine Learning and Computer Application (ICMLCA '23). Association for Computing Machinery, New York, NY, USA, 616–621. https://doi.org/10.1145/3650215.3650324
[14] Wu, H., Liu, X., Lei, J., & Wang, B. (2025). LSTM-Transformer based load interval prediction algorithm for governor oil pump. In 2025 37th Chinese Control and Decision Conference (CCDC). IEEE. https://doi.org/10.1109/CCDC65474.2025.11090619.
[15] Bastakoti, B., Parajuli, S., & Regmi, B. (2026). LSTM vs. Transformer Models in Power Forecasting: A Comprehensive Survey. Preprints. https://doi.org/10.20944/preprints202603.0624.v1.
[16] Tsai, Y.H., Jheng, Y.J., & Tsaih, R.H. (2019). The cramming, softening and integrating learning algorithm with parametric ReLU activation function for binary input/output problems. In 2019 International Joint Conference on Neural Networks (IJCNN) (pp. 1–8). IEEE. https://doi.org/10.1109/IJCNN.2019.8852023.
[17] Zoph, B., & Le, Q. V. (2017). Neural architecture search with reinforcement learning. In International Conference on Learning Representations (ICLR 2017). arXiv. https://arxiv.org/abs/1611.01578.
[18] Hong, T. & Fan, S. (2016). Probabilistic electric load forecasting: A tutorial review. International Journal of Forecasting, 32(3), 914–938. https://doi.org/10.1016/j.ijforecast.2015.11.011 .
[19] Putra, K. R. (2025). Comparison of prediction models: Decision tree, random forest, and support vector regression. Jurnal Informatika dan Rekayasa Perangkat Lunak (JATIKA), 6(1), 39–49. https://doi.org/10.33365/jatika.v6i1.18.
[20] Andrychowicz, M., Denil, M., Gomez, S., Hoffman, M. W., Pfau, D., Schaul, T., Shillingford, B., & de Freitas, N. (2016). Learning to learn by gradient descent by gradient descent. arXiv. https://doi.org/10.48550/arXiv.1606.04474
[21] Bischl, B., Binder, M., Lang, M., Pielok, T., Richter, J., Coors, S., Thomas, J., Ullmann, T., Becker, M., Boulesteix, A.-L., Deng, D., & Lindauer, M. (2023). Hyperparameter optimization: Foundations, algorithms, best practices, and open challenges. WIREs Data Mining and Knowledge Discovery, 13(2), Article e1484. https://doi.org/10.1002/widm.1484
[22] Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A. A., Veness, J., Bellemare, M. G., Graves, A., Riedmiller, M., Fidjeland, A. K., Ostrovski, G., Petersen, S., Beattie, C., Sadik, A., Antonoglou, I., King, H., Kumaran, D., Wierstra, D., Legg, S., & Hassabis, D. (2015). Human-level control through deep reinforcement learning. Nature, 518(7540), 529–533. https://doi.org/10.1038/nature14236.
[23] Schulman, J., Wolski, F., Dhariwal, P., Radford, A., & Klimov, O. (2017). Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347.
[24] Haarnoja, T., Zhou, A., Hartikainen, K., Tucker, G., Ha, S., Tan, J., Kumar, V., Zhu, H., Gupta, A., Abbeel, P., & Levine, S. (2018). Soft actor-critic algorithms and applications. arXiv preprint arXiv:1812.05905.
[25] Mnih, V., Badia, A. P., Mirza, M., Graves, A., Lillicrap, T., Harley, T., Silver, D., & Kavukcuoglu, K. (2016). Asynchronous methods for deep reinforcement learning. In Proceedings of the 33rd International Conference on Machine Learning (pp. 1928–1937).
[26] Marquardt, D. W. (1963). An algorithm for least-squares estimation of nonlinear parameters, SIAM Journal on Applied Mathematics, 11 , 431–441.
[27] Werbos, P. J. (1994). The Roots of Backpropagation: From Ordered Derivatives to Neural Networks and Political Forecasting . John Wiley & Sons, New York, NY, USA.
[28] Haykin, S. (2001). Neural networks: A comprehensive foundation. Pearson Education Press, New York, NY, USA.
[29] Tsaih, R. (1993). The softening learning procedure. Mathematical and Computer Modelling, 18(8), 61–64. https://doi.org/10.1016/0895-7177(93)90163-S.
[30] Tsaih, R.H. (1997). Reasoning neural networks. In S. W. Ellacott, J. C. Mason, & I. J. Anderson (Eds.), Mathematics of neural networks: Models, algorithms and applications (pp. 366–371). Springer.
[31] Tsaih, R. (1998). An explanation of reasoning neural networks. Mathematical and Computer Modelling, 28(2), 37–44. https://doi.org/10.1016/S0895-7177(98)00094-1.
[32] ESB. (2024). Ireland’s pathway to a net zero energy system by 2050. ESB.
[33] EirGrid. (2023, July 4). EirGrid revises its 2030 roadmap to reach 80% of renewable power. Enerdata. https://www.enerdata.net/publications/daily-energy-news/eirgrid-ireland-revises-its-2030-roadmap-reach-80-renewable-power.html.
[34] REGlobal. (2025, February 23). Ireland's renewable energy policy review. https://reglobal.org/irelands-renewable-energy-policy-review/ .
[35] KPMG Ireland. (2020). Act now: Accelerating onshore renewable energy in Ireland. KPMG. https://assets.kpmg.com/content/dam/kpmgsites/ie/pdf/insights/energy/ie-act-now-onshore-renewable-energy-ireland-3.pdf.
[36] Gomes, H.M., Bifet, A., Read, J., Barddal, J.P., Enembreck, F., Pfharinger, B., Holmes, G., & Abdessalem T. (2017). Adaptive random forests for evolving data stream classification. Mach Learn 106, 1469–1495. https://doi.org/10.1007/s10994-017-5642-8.
[37] Gregor, S., & Hevner, A. R. (2013). Positioning and presenting design science research for maximum impact. MIS Quarterly, 37(2), 337–355. https://doi.org/10.25300/MISQ/2013/37.2.01.
[38] Kumeno, F. (2019). Software engineering challenges for machine learning applications: A literature review. Intelligent Decision Technologies, 13(4), 463–476. https://doi.org/10.3233/IDT-190160.
[39] Benabbes, S., Aissaoui, W., & Boucetti, R. (2026). A lightweight edge-deployable ANN model for real-time energy anomaly detection in IoT-driven smart grids. Informatica, 50(1). https://doi.org/10.31449/inf.v50i1.12670.
[40] Liu, D., Zang, C., Zeng, P., Li, W., Wang, X., Liu, Y., Su, J., & Xu, S. (2023). Deep reinforcement learning for real-time economic energy management of microgrid system considering uncertainties. Frontiers in Energy Research, 11, Article 1163053. https://doi.org/10.3389/fenrg.2023.1163053.
[41] Qiu, D., Ye, Y., Papadaskalopoulos, D., & Strbac, G. (2020). A deep reinforcement learning method for pricing electric vehicles with discrete charging levels. IEEE Transactions on Industry Applications, 56(5), 5901–5912. https://doi.org/10.1109/TIA.2020.2984614.
[42] Liu, H., Sun, Y., & Wang, Y. (2025). A non-invasive load monitoring method for edge computing based on MobileNetV3 and dynamic time regulation. arXiv. https://doi.org/10.48550/arXiv.2504.16142.
[43] Moura, G. A., Soares, W. A., Belfort, D. R., & Catunda, S. Y. C. (2026). On-device edge AI for monthly energy forecasting on ESP32-S3 using 3-day windows: Hardware-constrained feasibility with harmonics and offline interharmonic benchmarking. IEEE Access, 14, 15414–15435. https://doi.org/10.1109/ACCESS.2025.3529471.
[44] Bonawitz, K., Eichner, H., Grieskamp, W., Huba, D., Ingerman, A., Ivanov, V., Kiddon, C., Konečný, J., Mazzocchi, S., McMahan, H. B., Van Overveldt, T., Petrou, D., Ramage, D., & Roselander, J. (2019). Towards federated learning at scale: System design. arXiv. https://doi.org/10.48550/arXiv.1902.01046.
[45] McMahan, B., Moore, E., Ramage, D., Hampson, S., & Agüera y Arcas, B. (2017). Communication-efficient learning of deep networks from decentralized data. In A. Singh & J. Zhu (Eds.), Proceedings of the 20th International Conference on Artificial Intelligence and Statistics (pp. 1273–1282). PMLR. https://proceedings.mlr.press/v54/mcmahan17a.html.
[46] Molokomme, D. N., Onumanyi, A. J., & Abu-Mahfouz, A. M. (2022). Edge intelligence in smart grids: A survey on architectures, offloading models, cyber security measures, and challenges. Journal of Sensor and Actuator Networks, 11(3), Article 47. https://doi.org/10.3390/jsan11030047.
[47] Zhou, Z., Chen, X., Li, E., Zeng, L., Luo, K., & Zhang, J. (2019). Edge intelligence: Paving the last mile of artificial intelligence with edge computing. Proceedings of the IEEE, 107(8), 1738–1762. https://doi.org/10.1109/JPROC.2019.2918951.
[48] Klaise, J., Van Looveren, A., Cox, C., Vacanti, G., & Coca, A. (2020). Monitoring and explainability of models in production. arXiv. https://arxiv.org/abs/2007.06299.
[49] Lundberg, S. M., & Lee, S.-I. (2017). A unified approach to interpreting model predictions. arXiv preprint arXiv:1705.07874. https://doi.org/10.48550/arXiv.1705.07874.