Introduction. The initial phase of research in this area focused on symbolic AI and knowledge-based systems. Early power grid forecasting relied on predefined models and expert systems that encoded domain knowledge about grid behavior, environmental factors, and operational rules. These systems typically used rule-based approaches, where expert insights were embedded to estimate grid performance under various conditions. However, their reliance on static knowledge bases made them inflexible in the face of unforeseen events. Moreover, they were computationally intensive and struggled with the complexity of large-scale grid operations and dynamic, real-time data. Despite these limitations, such approaches provided foundational insights into grid behavior and paved the way for more advanced methods. Despite the progress in data-driven forecasting, current methods still encounter limitations. These include the need for large volumes of high-quality data, difficulty in interpreting model outputs, and limited adaptability to rapidly changing grid conditions. To address these challenges, a novel approach is proposed that combines advanced deep learning with real-time data integration from diverse sources such as IoT sensors and weather forecasting models. This hybrid strategy improves forecasting accuracy by leveraging both long-term historical patterns and short-term real-time observations. It also incorporates adaptive algorithms that continuously update the model as new data becomes available, allowing the system to respond effectively to unexpected events and evolving grid dynamics [1], [2].

Purpose and objectives. Forecasting approaches range from statistical methods like autoregressive integrated moving average (ARIMA) to advanced machine learning algorithms such as artificial neural networks (ANN), support vector machines (SVM), and ensemble models. Recent advances in deep learning, including convolutional neural networks (CNN) and recurrent neural networks (RNN), have significantly improved accuracy by capturing complex temporal patterns and spatial variability in solar irradiance. Attention mechanisms and Transformer-based models are also being explored to enhance both performance and interpretability under volatile conditions. Accurate PV forecasting plays a critical role in strengthening grid resilience during disasters. Embedding PV forecasting within a broader analytics framework allows operators to better manage the risks of renewable energy variability.  The selection of base models in our Adaptive Ensemble Predictor (AEP) is grounded in the principle of maximizing diversity while ensuring complementary strengths across different data regimes. Decision trees, neural networks, linear regressors, and kernel machines are incorporated due to their distinct learning characteristics. Decision trees provide interpretable, rule-based modeling capabilities and are especially suitable for capturing abrupt transitions or decision thresholds -common in grid protection logic and load-shedding conditions. However, they tend to overfit and have high variance in noisy or continuous-valued datasets. Neural networks, including CNNs and LSTMs, are capable of capturing complex nonlinear patterns and temporal dependencies. These architectures are particularly effective in processing spatiotemporal features from PV generation and sensor telemetry, but may suffer from instability in sparse or adversarial inputs [3].

The diagram, Fig. 1, illustrates the integration of heterogeneous base modeling, positional and user embeddings, and context embedding within the AEP framework. It showcases how multiple specialized models, including neural networks, transformers, and base learners, are combined to enhance predictive performance and generalization. The adaptive weighting mechanism adjusts the contribution of each model based on its validation performance, promoting robust learning across diverse time series data  [4].

Figure 1. Schematic diagram of the Heterogeneous Base Modeling

Linear models and kernel methods add valuable low-bias estimators that are robust in small-data or high-noise settings, offering regularized decision functions with strong generalization. By combining these models, the ensemble benefits from both expressiveness and robustness. The adaptive weighting scheme ensures that models with stronger validation performance are prioritized without excluding weaker learners completely, thus balancing diversity and accuracy. The Adaptive Ensemble Predictor (AEP) distinguishes itself through a performance-based weighting scheme that dynamically calibrates the contribution of each base learner according to its validation performance. Unlike static averaging strategies or uniform weight allocations. While individual components of ensemble – such as decision trees and neural networks – are commonly used in machine learning, the novelty of our method lies in how they are integrated and optimized jointly for disaster-resilient forecasting. Unlike conventional deep learning models (e.g., LSTM, Transformer), our framework adopts a performance ware adaptive weighting mechanism that dynamically adjusts model contributions based on validation error. Furthermore, the ensemble is trained via joint end-to-end optimization, rather than in separate stages, enabling tighter coordination across diverse learners. The model also incorporates multi-stage feature reuse and hard example mining, which improve generalization and stability under volatile inputs. These innovations are particularly suitable for the power grid context, where real-time adaptability, interpretability, and robustness are critical. Therefore, TSPNet offers a structured yet flexible alternative to deep homogeneous architectures, enhancing both modeling performance and operational usability in disaster scenarios [5].

Figure 2. Schematic diagram of the unified Adaptive Learning Strategy

The diagram, Fig. 2, illustrates the components of the proposed model, highlighting the integration of adaptive ensemble weighting, feature map F.L.M blocks, multi-stage learning, and hard example mining. Key elements such as residual connections, down-sampling, and up-sampling operations are shown, demonstrating how the model processes input data at multiple stages to refine predictions and improve generalization. The combination of these mechanisms allows the model to dynamically adjust its focus and enhance performance across diverse tasks [6], [7].

The diagram, Fig. 3, illustrates the workflow of identifying and prioritizing hard examples during training. It shows how the model dynamically adjusts its focus on difficult-to-predict instances by assigning higher weights to them, thereby refining decision boundaries and improving model robustness. Key components include the residual block, the computation of prediction confidence, and the iterative process of updating the hard example set based on previous stage residuals [8], [9].

Figure 3. Schematic diagram of the Learning from Hard Examples

The adaptive ensemble weights are optimized using a softmax-based reweighting scheme that dynamically adjusts according to the mean squared error (MSE) on a validation set. The models were evaluated on four benchmark datasets with varying temporal and structural properties:

1. TimeTravel Competition Dataset: A heterogeneous time series benchmark containing industrial and environmental monitoring signals.

2. UTSD Dataset: A transportation system dataset featuring real-time traffic flow, vehicle speeds, and occupancy metrics from urban sensors.

3. TimeHetNet Dataset: A temporally heterogeneous network dataset involving user behaviors, interactions, and multitype event dynamics.

4. OpenEI Dataset: A smart grid dataset including power generation, consumption, and market pricing from distributed energy systems.

Figure 4. Comparison of Time Series Prediction Models on TimeTravel Competition and UTSD Datasets

Traditional grid forecasting models, which often rely on rule-based systems and limited datasets, struggle to capture the dynamic and nonlinear interactions between renewable energy sources like PV plants and the power grid, especially under high-stress conditions such as natural disasters. These conventional methods are often insufficient in terms of scalability, accuracy, and adaptability in real-time disaster scenarios. To overcome these shortcomings, the authors propose a novel framework that incorporates machine learning techniques to improve forecasting accuracy and enhance grid resilience. By using a combination of real-time data from multiple sources – such as smart meters, environmental sensors, weather services, and historical logs – and advanced algorithms including decision trees, artificial neural networks (ANN), and ensemble methods like random forests or gradient boosting, the proposed approach enables robust and precise power predictions. These predictions, when fed into grid control and management systems, improve operational readiness and real-time disaster response, optimizing resource allocation, load balancing, and reducing downtime. Experimental results demonstrate that the proposed method outperforms traditional approaches in terms of forecasting accuracy, reactivity, and adaptive learning under fluctuating conditions, offering a promising direction to modernize grid operations and strengthen infrastructure resilience [10].

In practical terms, TSPNet offers its most significant advantages under operational conditions characterized by volatility, disruption, or limited computational infrastructure. It outperforms simpler or traditional approaches when forecasting tasks require integration of noisy, missing, or real-time heterogeneous data sources – as is often the case during power grid disasters. In contrast to static models such as ARIMA, or homogeneous architectures like LSTM and Transformer, TSPNet dynamically adapts to changing data conditions through performancebased ensemble learning and hard example mining. Moreover, it remains viable in resource-constrained environments due to its efficient modular structure. While TSPNet may not universally dominate in static, well-instrumented grid settings, its strengths are most evident in uncertain, highrisk scenarios where traditional methods degrade. This positioning defines both its scope and value within the broader landscape of grid forecasting solutions.

References:

1. X. Wen and W. Li. Time series prediction based on lstm-attention-lstm model, IEEE Access, 2023.

2. L. Ren, Z. Jia, Y. Laili, and D.-W. Huang. Deep learning for time-series prediction in iiot: Progress, challenges, and prospects, IEEE Transactions on Neural Networks and Learning Systems, 2023.

3. Y. Li, K. Wu, and J. Liu. Self-paced arima for robust time series prediction, Knowledge-Based Systems, 2023.

4. L. Yin, L. Wang, T. Li, S. Lu, J. Tian, Z. Yin, X. Li, and W. Zheng. Unetlstm: Time series-enhanced lake boundary prediction model, Land, 2023.

5. C. Yu, F. Wang, Z. Shao, T. Sun, L. Wu, and Y. Xu. Dsformer: A double sampling transformer for multivariate time series long-term prediction, International Conference on Information and Knowledge Management, 2023.

6. D. M. Durairaj and B. G. K. Mohan. A convolutional neural network based approach to financial time series prediction, Neural computing &applications (Print), 2022.

7. H. V. Dudukcu, M. Taskiran, Z. G. C. Taskiran, and T. Yıldırım. Temporal convolutional networks with rnn approach for chaotic time series prediction, Applied Soft Computing, 2022.

8. Y. Xiao, H. Yin, Y. Zhang, H. Qi, Y. Zhang, and Z. Liu. A dual-stage attention-based conv-lstm network for spatio-temporal correlation and multivariate time series prediction, International Journal of Intelligent Systems, 2021.

9. X. Wen and W. Li. Time series prediction based on lstm-attention-lstm model, IEEE Access, vol. 11, pp. 130, 456–130, 469, 2023.

10. Z. Karevan and J. A. K. Suykens, “Transductive lstm for time-series prediction: An application to weather forecasting”, Neural Networks, vol. 125, pp. 1–9, 2020.