Résultats de recherche pour « Temporal Fusion Transformer »

Bu çalışmada, derin öğrenme modellerinden biri olan Temporal Fusion Transformer (TFT) mimarisi kullanılarak Türkiye’ye ait aylık İnşaat Maliyet Endeksi (İME) öngörüsü gerçekleştirilmeye çalışılmıştır. Burada, günümüzde öngörüde kullanılan derin öğrenme modellerinden biri olan TFT modelini tanıtmak ve inşaat maliyet endeksi öngörüsünde kullanarak performansını görmek amaçlanmıştır. İME öngörüsü için, Ocak 2015-Eylül 2025 zaman aralığında aylık veriler kullanılmıştır. Ayrıca, maliyet dinamiklerini etkileyebileceği değerlendirilen üç bağımsız değişken (inşaat yapı sayısı (), sanayi üretim endeksi, konut satış endeksi) seçilerek modele dâhil edilmiştir. Öngörüde TFT modeli kullanılırken, model performansı Kök Ortalama Kare Hata (RMSE) metriği ile değerlendirilmiştir. TFT modelinin RMSE değeri 36,01 olarak bulunmuştur. Elde edilen bulgular, TFT modelinin hem kısa vadeli öngörülerde hem de tüm dönem performansında yüksek doğruluk sağladığını ve trend ile volatilite dinamiklerini başarılı biçimde yakaladığını göstermektedir. Ayrıca değişken önem analizi sayesinde, inşaat maliyetlerini etkileyen temel göstergelerin göreli katkıları nicel olarak ortaya konulmuştur.

Modeling the complex spatio-temporal dynamics of infectious diseases presents a significant computational challenge due to heterogeneous regional interactions, high-dimensional multimodal data streams, and the critical need for algorithmic fairness. This paper proposes a novel computational framework that unifies graph-based spatio-temporal forecasting, anomaly detection, and retrieval-augmented generation (RAG) into a single mathematical architecture. The predictive backbone employs a graph-augmented Temporal Fusion Transformer to capture non-linear temporal dependencies and spatial disease propagation. By formalizing regional topology and mobility flows as a weighted mathematical graph, the model systematically integrates structured epidemiological counts, continuous environmental covariates, and digital trace signals. To address algorithmic bias, we formulate a fairness-aware optimization problem by embedding a specific regularization term into the training objective, which mathematically penalizes disparities in true-positive rates across diverse socio-demographic strata. Furthermore, the numerical outputs and anomaly scores are processed by a large language model equipped with hybrid dense and sparse retrieval to generate interpretable, computationally grounded decision support. Extensive experiments on a longitudinal dataset comprising 62 administrative regions over 260 weeks validate the mathematical robustness of the proposed framework. The graph-augmented architecture improved forecasting accuracy by up to 24% and anomaly detection F1 scores by over 6% compared to state-of-the-art deep learning baselines, while the fairness-regularized loss function reduced the maximum subgroup recall gap by more than 50%. These findings demonstrate that predictive accuracy and algorithmic fairness can be jointly optimized, providing a rigorous computational methodology for spatio-temporal epidemic modeling and AI-driven surveillance.

Accurately forecasting lithium carbonate prices is essential for strategic planning in the global energy-storage and electric-vehicle industries. This study proposes a Temporal Fusion Transformer (TFT)–based forecasting framework to model daily lithium market dynamics using multi-source historical data from January 2012 to September 2025. The model leverages attention mechanisms, temporal gating, and variable-importance learning. Using these learned dynamics, the TFT model generates long-term forecasts for 2026–2030, producing smooth, stable, and economically plausible price trajectories. Its performance is compared with Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) baselines, where absolute-error box plots and numerical performance metrics demonstrate that the TFT model yields substantially lower variance, reduced forecast deviation, and more realistic long-horizon behavior. From the perspective of explainable artificial intelligence (XAI), the research findings provide a diverse range of stakeholders in lithium-dependent industries with reliable, transparent, and interpretable insights, thereby effectively supporting their long-term strategic decision-making processes regarding critical raw material price forecasts.

Abstract Urban traffic flow prediction is critical for intelligent transportation systems, yet remains challenging due to complex spatiotemporal dependencies and external weather influences. Existing methods suffer from three key limitations: (1) uniform time-domain modeling fails to disentangle periodic trends from aperiodic fluctuations, (2) weather information is integrated through simple concatenation without considering multi-granularity impacts, and (3) lack of interpretability hinders understanding of weather-traffic interactions. To address these issues, we propose MFTFormer, a novel framework that synergistically combines time-frequency decomposition with hierarchical weather-aware fusion. Our approach introduces three key innovations: (1) a dual-domain encoder with mixture-of-experts that separately processes trend and seasonal components via specialized frequency and temporal attention networks, (2) a block-aligned fusion module that hierarchically reorganizes features through convolutional refinement and weather-conditioned modulation, capturing both immediate and cumulative meteorological effects, and (3) interpretable attention mechanisms that reveal dynamic spatial dependencies under varying weather conditions. Extensive experiments on four real-world datasets (PeMS04, PeMS08, Metro-Traffic, METR-LA) demonstrate that MFTFormer achieves state-of-the-art performance with an average 6.2% MAE reduction over strong baselines, while providing actionable insights into weather-traffic relationships through attention visualization.

Abstract Type 1 Diabetes Mellitus (T1DM) is a chronic autoimmune condition characterized by absolute insulin deficiency that affects an estimated 9.1 million individuals worldwide. Hypoglycemia is still a barrier to safe glycemic control, and existing prediction tools have high false alarm rates and limited integration of multimodal physiological signals. We present an attention-augmented Temporal Fusion Transformer (TFT) trained on the T1DiabetesGranada longitudinal dataset (73 adults) for 30 minute horizon hypoglycemia forecasting. We use continuous glucose monitoring (CGM), forward filled biochemical markers, and demographic covariates which allows the model to achieve a 5 split cross validated AUC of 0.921±0.005, significantly outperforming a glucose only LSTM baseline (AUC 0.853±0.007; mean difference 0.068, 95% CI [0.063, 0.073], paired t -test p = 3.4 × 10 -6 ) and a glucose only vanilla Transformer (single split AUC 0.837). The TFT also reduces false alarm rate by around 36% relative to the LSTM baseline and improves sensitivity. An ablation study shows that removing the temporal convolutional block or the static covariate pathway produces negligible change in AUC, meaning that in this configuration the LSTM attention backbone already gets most of the predictive signal; we talk about this finding and its implications for feature engineering in future work. These results show that multimodal, attention-based sequence models can provide more sensitive and lower latency hypoglycemia risk estimates than glucose only or rule based options, which will lead to further work toward prospective validation.

Reliable runoff forecasting is critical for water management and flood preparedness in Nepal’s steep, data-scarce catchments. Traditional models such as SWAT provide process insights but demand extensive calibration and detailed inputs often unavailable in such regions. Recent advances in attentionbased deep learning offer new opportunities to capture temporal dependencies with improved interpretability. This study evaluates the Temporal Fusion Transformer (TFT) for monthly runoff prediction using 40 years (1980–2020) of hydrometeorological data from Nepal, benchmarked against Random Forest (RF) and Long Short-Term Memory (LSTM) networks. Results show that RF underestimates peaks, LSTM captures seasonality but falters under monsoon extremes, while TFT consistently achieves superior accuracy (RMSE = 22.5, R2 = 0.88). Attention weights further reveal precipitation and antecedent runoff as dominant drivers, reinforcing hydrological understanding. These findings highlight attention-based architectures as accurate and interpretable tools for operational flood forecasting and climate-resilient water management.

Accurate traffic prediction is crucial for optimizing taxi demand, managing traffic flow, and planning public transportation routes. Traditional models often fail to capture complex spatial–temporal dependencies. To tackle this, we introduce the Spatio-Temporal Embedding Fusion Transformer (STEFT). This deep learning model leverages attention mechanisms and feature fusion to effectively model dynamic dependencies in traffic data. STEFT includes an Embedding Fusion Network that integrates spatial, temporal, and flow embeddings, preserving original flow information. The Flow Block uses an enhanced Transformer encoder to capture periodic dependencies within neighboring regions, while the Prediction Block forecasts inflow and outflow dynamics using a fully connected network. Experiments on NYC (New York City) Taxi and NYC Bike datasets show STEFT’s superior performance over baseline methods in RMSE and MAPE metrics, highlighting the effectiveness of the concatenation-based feature fusion approach. Ablation studies confirm the contribution of each component, underscoring STEFT’s potential for real-world traffic prediction and other spatial–temporal challenges.

Abstract Accurate real-time forecasting of remaining vehicle range remains a challenge, particularly under varying payloads, road gradients, and environmental conditions. Conventional Temporal Fusion Transformers (TFTs) utilize attention mechanisms to dynamically weight historical inputs but may fail to capture explicit physical relationships that are critical for accurate predictions in heavy-duty electric trucks. This paper introduces a novel approach to integrating physical vehicle dynamics directly into the attention mechanism of TFTs. Our method, Physics-Informed Attention for TFT (PIA-TFT), modifies attention calculation by injecting physics-based relevance scores derived from vehicle speed, payload, road gradients, and other physical parameters, improving interpretability and model accuracy under operational conditions. Empirical evaluations conducted with real-world data from electric trucks demonstrate that the PIA-TFT reduces prediction errors compared to standard TFTs by up to 18%. Our approach is a step towards more physically consistent and explainable deep learning architectures for automotive forecasting tasks.

Abstract Accurate traffic forecasting is challenging due to the intricate inter-dependencies of road networks and congestion caused by unexpected accidents. Recent work has focused on dynamically changing traffic characteristics but has paid less attention to the global cross-spatial-temporal domain of modeling, which may limit their performance. In this paper, we propose a novel plug-and-play fusion unit to accurately express the spatial-temporal dependencies by cross-domain complementary information integration, named the Cross-Domain Transformer Spatial-Temporal Fusion Network (CDTSTFN). By introducing two-stage fusion units, we compensate information loss and resolve the mismatch in fused information. This enables CDTSTFN to largely augment the base spatial-temporal predictors with learned both local-global spatial and short-long temporal dependencies on cross-domain spatial-temporal patterns. A comprehensive set of both quantitative and qualitative assessments is performed on six public traffic network datasets (PeMS03, PeMS04, PeMS07, PeMS08, METR-LA, and PeMS-BAY), demonstrating the superior performance of our model.