demand forecasting

This paper addresses the critical issue of prediction instability in time-series forecasting models for supply chain demand planning—where identical inputs yield highly variable outputs, increasing manual intervention and eroding planner trust. Departing from conventional accuracy-centric evaluation paradigms, we formally define and prioritize *prediction stability*—quantified as the variance of repeated forecasts over fixed inputs—as a key operational deployment metric, and systematically investigate how intrinsic model stochasticity undermines output consistency. Experiments on the M5 and Favorita benchmarks compare state-of-the-art models including Chronos, DeepAR, PatchTST, TFT, TiDE, and AutoGluon. Results demonstrate that ensemble-based approaches significantly reduce forecast variance while preserving accuracy, thereby enhancing stability. Our work establishes stability as a first-class criterion for model selection and deployment in production, advancing the field from an “accuracy-first” to a “stability-accuracy balanced” paradigm.

This study addresses the limitation of traditional demand forecasting models that rely on statistical metrics such as MAE and RMSE, which often fail to capture their real-world impact on inventory key performance indicators (KPIs)—particularly total cost and service level—in intermittent demand contexts like automotive aftermarket spare parts. To bridge this gap, the authors propose a decision-centric simulation framework that integrates a synthetic demand generator, a plug-and-play forecasting module, and an inventory control simulator. This framework systematically establishes, for the first time, a mapping between forecast errors and inventory KPIs, enabling end-to-end evaluation of any forecasting model under realistic inventory policies. It reveals that improvements in statistical accuracy do not necessarily translate into better operational performance and provides a cost–service trade-off–oriented basis for model selection.

Time-series forecasting suffers from poor generalization and low efficiency due to tight coupling among sequence representation, information extraction, and future projection. To address this, we propose a modular forecasting framework that decouples the pipeline into three independently optimizable stages—representation learning, information extraction, and target projection—enabling flexible, task-aware component configuration. Our approach innovatively integrates convolutional layers with a lightweight self-attention mechanism, achieving efficient local feature modeling while capturing long-range temporal dependencies. Evaluated on seven benchmark datasets, the method consistently outperforms existing state-of-the-art models in prediction accuracy, while requiring significantly fewer parameters and achieving faster training and inference speeds. This demonstrates substantial improvements in both statistical performance and computational efficiency.

To address the high computational overhead and poor scalability caused by frequent retraining of ML-based capacity prediction models in capacity management, this paper proposes an on-demand retraining mechanism driven by data drift detection. The method integrates time-series analysis, lightweight ML forecasting models, and multi-dimensional data drift detection within an AIOps framework to enable dynamic capacity prediction. Its core innovation lies in triggering retraining only upon statistically significant data drift—eliminating redundant periodic updates. Experiments show that, under most dynamic workloads, its prediction accuracy matches that of periodic retraining (error difference <3%), while reducing average computational cost by 62%. Only under extremely high-frequency workload shifts is periodic retraining recommended as a fallback. This work establishes a new paradigm for resource demand forecasting that jointly optimizes accuracy, efficiency, and adaptability.

Accurate multi-level electricity load forecasting is critical for maintaining supply–demand balance in smart grids. This work constructs a unified benchmark dataset spanning transmission system operators (TSOs), medium-voltage feeders, and low-voltage consumers, and systematically evaluates ten short-term load forecasting methods. The study introduces YAformer, a flexible and scalable Transformer-based architecture, and demonstrates—for the first time under a consistent benchmark—the general superiority of Transformers in multi-scale load forecasting. Standard Transformer models outperform most complex variants, with long input contexts, effective covariate integration, and continual retraining identified as key performance drivers. Experimental results show that Transformer-based approaches reduce prediction errors by 6.6%–10.7%, while Chronos-2, though effective in zero-shot settings, struggles to capture TSO-level exceptional events.

This study addresses the challenge of inaccurate prediction intervals in smart building load forecasting, often caused by missing inputs and reconstruction errors, which undermines demand response effectiveness. The authors propose a unified day-ahead probabilistic forecasting framework that integrates temporal alignment, missing input reconstruction, and causal feature extraction to systematically compare two uncertainty modeling strategies: modular post-processing quantile regression and in-model quantile learning. Experiments across three medium-scale backbone architectures—RNN, hybrid RNN, and Temporal Fusion Transformer (TFT)—demonstrate that TFT with embedded quantile learning yields more reliable prediction intervals, achieving approximately five times narrower widths and better coverage than modular approaches. Although input reconstruction increases the quantile score by 106%, it leaves interval width largely unchanged, revealing inherent limitations of post-processing methods under reconstruction scenarios.

This study addresses the challenge of predicting individual task labor demand under variable forecasting horizons caused by heterogeneous task durations in construction projects, where predictions must adhere to a predefined total labor constraint. To this end, the authors propose the Constraint-Preserving Residual Allocation Forecasting (CP-RAF) method, which encodes historical labor sequences into temporal shape coefficient vectors and generates a labor distribution profile for the remaining duration by retrieving and similarity-weighting completed tasks. This profile dynamically allocates the total labor quota and adjusts the prediction horizon accordingly. CP-RAF is the first approach to explicitly embed the total labor constraint directly into the forecasting process, balancing operational feasibility with predictive accuracy. Experimental results on real-world construction site data demonstrate that CP-RAF significantly outperforms eight baseline models, achieving consistently low prediction errors across both medium- and long-term forecasting scenarios with fixed and variable horizons.