High-quality wind power forecasting is crucial for the operation of modern power grids. However, prevailing data-driven paradigms either train a site-specific model which cannot generalize to other locations or rely on fine-tuning of general-purpose time series foundation models which are difficult to incorporate domain-specific data in the energy sector. This paper introduces WindFM, a lightweight and generative Foundation Model designed specifically for probabilistic wind power forecasting. WindFM employs a discretize-and-generate framework. A specialized time-series tokenizer first converts continuous multivariate observations into discrete, hierarchical tokens. Subsequently, a decoder-only Transformer learns a universal representation of wind generation dynamics by autoregressively pre-training on these token sequences. Using the comprehensive WIND Toolkit dataset comprising approximately 150 billion time steps from more than 126,000 sites, WindFM develops a foundational understanding of the complex interplay between atmospheric conditions and power output. Extensive experiments demonstrate that our compact 8.1M parameter model achieves state-of-the-art zero-shot performance on both deterministic and probabilistic tasks, outperforming specialized models and larger foundation models without any fine-tuning. In particular, WindFM exhibits strong adaptiveness under out-of-distribution data from a different continent, demonstrating the robustness and transferability of its learned representations. Our pre-trained model is publicly available at https://github.com/shiyu-coder/WindFM.

Autonomous Mobile Robots (AMRs) operate on battery power, making energy efficiency a critical consideration particularly in outdoor environments where terrain variations affect energy consumption. While prior research has primarily focused on computing energy-efficient paths from a source to a destination, these approaches often overlook practical scenarios where a robot needs to pick up an object en route--an action that can significantly impact energy consumption due to changes in payload. This paper introduces the Object-Pickup Minimum Energy Path Problem (OMEPP), which addresses energy-efficient route planning for Autonomous Mobile Robots (AMRs) required to pick up an object from one of the many possible locations and take it to a destination. To address the OMEPP problem, we first introduce a baseline algorithm that employs the Z* algorithm, a variant of A* tailored for energy-efficient routing, to iteratively visit each pickup point. While this approach guarantees optimality, it suffers from high computational cost due to repeated search efforts at each pickup location. To mitigate this inefficiency, we propose a concurrent PCPD search that manages multiple Z* searches simultaneously across all pickup points. Central to our solution is the Payload-Constrained Path Database (PCPD), an extension of the Compressed Path Database (CPD), a state-of-the-art technique for fast shortest path computation, that incorporates payload constraints. We further demonstrate that PCPD significantly reduces branching factors during search, leading to improved overall performance. Although the concurrent PCPD search may produce slightly suboptimal solutions, extensive experiments on real-world datasets demonstrate that it achieves near-optimal performance while being one to two orders of magnitude faster than the baseline algorithm derived from existing methods.

While queueing network models are powerful tools for analyzing service systems, they traditionally require substantial human effort and domain expertise to construct. To make this modeling approach more scalable and accessible, we propose a data-driven framework for queueing network modeling and simulation based on autoregressive sequence models trained on event-stream data. Instead of explicitly specifying arrival processes, service mechanisms, or routing logic, our approach learns the conditional distributions of event types and event times, recasting the modeling task as a problem of sequence distribution learning. We show that Transformer-style architectures can effectively parameterize these distributions, enabling automated construction of high-fidelity simulators. As a proof of concept, we validate our framework on event tables generated from diverse queueing networks, showcasing its utility in simulation, uncertainty quantification, and counterfactual evaluation. Leveraging advances in artificial intelligence and the growing availability of data, our framework takes a step toward more automated, data-driven modeling pipelines to support broader adoption of queueing network models across service domains.

Spatio-temporal forecasting aims to predict the future state of dynamic systems and plays an important role in multiple fields. However, existing solutions only focus on modeling using a limited number of observed target variables. In real-world scenarios, exogenous variables can be integrated into the model as additional input features and associated with the target signal to promote forecast accuracy. Although promising, this still encounters two challenges: the inconsistent effects of different exogenous variables to the target system, and the imbalance effects between historical variables and future variables. To address these challenges, this paper introduces \model, a novel framework for modeling \underline{exo}genous variables in \underline{s}patio-\underline{t}emporal forecasting, which follows a ``select, then balance'' paradigm. Specifically, we first construct a latent space gated expert module, where fused exogenous information is projected into a latent space to dynamically select and recompose salient signals via specialized sub-experts. Furthermore, we design a siamese network architecture in which recomposed representations of past and future exogenous variables are fed into dual-branch spatio-temporal backbones to capture dynamic patterns. The outputs are integrated through a context-aware weighting mechanism to achieve dynamic balance during the modeling process. Extensive experiments on real-world datasets demonstrate the effectiveness, generality, robustness, and efficiency of our proposed framework.

Decision-making under uncertainty in energy management is complicated by unknown parameters hindering optimal strategies, particularly in Battery Energy Storage System (BESS) operations. Predict-Then-Optimise (PTO) approaches treat forecasting and optimisation as separate processes, allowing prediction errors to cascade into suboptimal decisions as models minimise forecasting errors rather than optimising downstream tasks. The emerging Decision-Focused Learning (DFL) methods overcome this limitation by integrating prediction and optimisation; however, they are relatively new and have been tested primarily on synthetic datasets or small-scale problems, with limited evidence of their practical viability. Real-world BESS applications present additional challenges, including greater variability and data scarcity due to collection constraints and operational limitations. Because of these challenges, this work leverages Automated Feature Engineering (AFE) to extract richer representations and improve the nascent approach of DFL. We propose an AFE-DFL framework suitable for small datasets that forecasts electricity prices and demand while optimising BESS operations to minimise costs. We validate its effectiveness on a novel real-world UK property dataset. The evaluation compares DFL methods against PTO, with and without AFE. The results show that, on average, DFL yields lower operating costs than PTO and adding AFE further improves the performance of DFL methods by 22.9-56.5% compared to the same models without AFE. These findings provide empirical evidence for DFL's practical viability in real-world settings, indicating that domain-specific AFE enhances DFL and reduces reliance on domain expertise for BESS optimisation, yielding economic benefits with broader implications for energy management systems facing similar challenges.

Energy forecasting is vital for grid reliability and operational efficiency. Although recent advances in time series forecasting have led to progress, existing benchmarks remain limited in spatial and temporal scope and lack multi-energy features. This raises concerns about their reliability and applicability in real-world deployment. To address this, we present the Real-E dataset, covering over 74 power stations across 30+ European countries over a 10-year span with rich metadata. Using Real- E, we conduct an extensive data analysis and benchmark over 20 baselines across various model types. We introduce a new metric to quantify shifts in correlation structures and show that existing methods struggle on our dataset, which exhibits more complex and non-stationary correlation dynamics. Our findings highlight key limitations of current methods and offer a strong empirical basis for building more robust forecasting models

Built on Llama 3.1-8B and scaled to 10B parameters, Llama-GENBA-10B is continuously pretrained on 164B tokens (82B English, 82B German, and 80M Bavarian), balancing resources while preventing English dominance. Targeted at the German NLP community, the model also promotes Bavarian as a low-resource language. Development tackled four challenges: (1) curating a multilingual corpus despite Bavarian scarcity, (2) creating a unified tokenizer for English, German, and Bavarian, (3) optimizing architecture and language-ratio hyperparame-ters for cross-lingual transfer, and (4) establishing the first standardized trilingual evaluation suite by translating German benchmarks into Bavarian. Evaluations show that Llama-GENBA-10B achieves strong cross-lingual performance, with the fine-tuned variant surpassing Apertus-8B-2509 and gemma-2-9b in Bavarian and establishing itself as the best model in its class for this language, while also outperforming EuroLLM in English and matching its results in German. Training on the Cerebras CS-2 demonstrated efficient large-scale multilingual pretraining with documented energy use, offering a blueprint for inclusive foundation models that integrate low-resource languages.

Mars exploration requires precise and reliable terrain models to ensure safe rover navigation across its unpredictable and often hazardous landscapes. Stereoscopic vision serves a critical role in the rover's perception, allowing scene reconstruction by generating precise depth maps through stereo matching. State-of-the-art Martian planetary exploration uses traditional local block-matching, aggregates cost over square windows, and refines disparities via smoothness constraints. However, this method often struggles with low-texture images, occlusion, and repetitive patterns because it considers only limited neighbouring pixels and lacks a wider understanding of scene context. This paper uses Semi-Global Matching (SGM) with superpixel-based refinement to mitigate the inherent block artefacts and recover lost details. The approach balances the efficiency and accuracy of SGM and adds context-aware segmentation to support more coherent depth inference. The proposed method has been evaluated in three datasets with successful results: In a Mars analogue, the terrain maps obtained show improved structural consistency, particularly in sloped or occlusion-prone regions. Large gaps behind rocks, which are common in raw disparity outputs, are reduced, and surface details like small rocks and edges are captured more accurately. Another two datasets, evaluated to test the method's general robustness and adaptability, show more precise disparity maps and more consistent terrain models, better suited for the demands of autonomous navigation on Mars, and competitive accuracy across both non-occluded and full-image error metrics. This paper outlines the entire terrain modelling process, from finding corresponding features to generating the final 2D navigation maps, offering a complete pipeline suitable for integration in future planetary exploration missions.

The You Only Look Once (YOLO) architecture is crucial for real-time object detection. However, deploying it in resource-constrained environments such as unmanned aerial vehicles (UAVs) requires efficient transfer learning. Although layer freezing is a common technique, the specific impact of various freezing configurations on contemporary YOLOv8 and YOLOv10 architectures remains unexplored, particularly with regard to the interplay between freezing depth, dataset characteristics, and training dynamics. This research addresses this gap by presenting a detailed analysis of layer-freezing strategies. We systematically investigate multiple freezing configurations across YOLOv8 and YOLOv10 variants using four challenging datasets that represent critical infrastructure monitoring. Our methodology integrates a gradient behavior analysis (L2 norm) and visual explanations (Grad-CAM) to provide deeper insights into training dynamics under different freezing strategies. Our results reveal that there is no universal optimal freezing strategy but, rather, one that depends on the properties of the data. For example, freezing the backbone is effective for preserving general-purpose features, while a shallower freeze is better suited to handling extreme class imbalance. These configurations reduce graphics processing unit (GPU) memory consumption by up to 28% compared to full fine-tuning and, in some cases, achieve mean average precision (mAP@50) scores that surpass those of full fine-tuning. Gradient analysis corroborates these findings, showing distinct convergence patterns for moderately frozen models. Ultimately, this work provides empirical findings and practical guidelines for selecting freezing strategies. It offers a practical, evidence-based approach to balanced transfer learning for object detection in scenarios with limited resources.

The enormous energy demand of artificial intelligence is driving the development of alternative hardware for deep learning. Physical neural networks try to exploit physical systems to perform machine learning more efficiently. In particular, optical systems can calculate with light using negligible energy. While their computational capabilities were long limited by the linearity of optical materials, nonlinear computations have recently been demonstrated through modified input encoding. Despite this breakthrough, our inability to determine if physical neural networks can learn arbitrary relationships between data -- a key requirement for deep learning known as universality -- hinders further progress. Here we present a fundamental theorem that establishes a universality condition for physical neural networks. It provides a powerful mathematical criterion that imposes device constraints, detailing how inputs should be encoded in the tunable parameters of the physical system. Based on this result, we propose a scalable architecture using free-space optics that is provably universal and achieves high accuracy on image classification tasks. Further, by combining the theorem with temporal multiplexing, we present a route to potentially huge effective system sizes in highly practical but poorly scalable on-chip photonic devices. Our theorem and scaling methods apply beyond optical systems and inform the design of a wide class of universal, energy-efficient physical neural networks, justifying further efforts in their development.