The core challenge in time series forecasting lies in effectively modeling long-term dependencies and multi-scale patterns. Although the NHiTS model has made progress in long-term forecasting through its multi-scale framework, its core multilayer perceptron (MLP) building blocks have limitations in feature representation capability, making it difficult to jointly capture local fine-grained patterns and global longterm dependencies. To address this, this paper proposes an improved model architecture, with its core innovation being the design of a novel Multi-Scale Fusion Block (MSFB) to enhance multi-period feature representation. This module explicitly models local temporal patterns and global dependencies through parallel multi-scale 1D convolutions and block-sparse attention mechanisms, respectively, and introduces a learnable dynamic fusion gating mechanism to adaptively integrate heterogeneous feature streams. Experiments were conducted on four benchmark datasets—ETTm2, Traffic, Weather, and Exchange—for training, validation, and testing. The results show that the improved model achieved average reductions of 11.20% and 7.79% in MAE and MSE metrics, respectively, compared to the original NHiTS model, and significantly outperformed mainstream comparative models such as TimesNet, Autoformer, and FEDformer. This validates the effectiveness of the proposed module in enhancing temporal representation learning and improving forecasting accuracy.

Tiancai Zhu, Jiangtao Liu, Jiahao Duan, Wei Wang, Yonghao Wu, Tongzhu Zhao. NHiTS-MSFB: Dynamic Local–Global Feature Fusion for Time Series Forecasting. Academic Journal of Computing & Information Science (2026), Vol. 9, Issue 1: 48-55. https://doi.org/10.25236/AJCIS.2026.090106.

[1] Li Yunfeng, Cai Ziwen, Zhao Yun, et al. Improved time series network for short-term power load forecasting method [J]. Journal of Hunan University (Natural Sciences), 2025, 52(10):205-216.

[2] Li Yong, Wu Hanxin, Li Chanxiao, et al. Short-term load forecasting for distribution network considering multi-dimensional meteorological indicators [J]. Smart Power, 2025, 53(06):116-123.

[3] Zhu Xiaotong, Lin Peiguang, Sun Mei, et al. Multivariate time series forecasting of financial data based on APDFinformer model [J]. Journal of Nanjing University (Natural Science), 2024, 60(06):930-939. 

[4] Wang C, Chen Y, Zhang S, et al. Stock market index prediction using deep Transformer model[J]. Expert Systems with Applications, 2022, 208: 118128.

[5] Fan J, Zhang K, Huang Y, et al. Parallel spatio-temporal attention-based TCN for multivariate time series prediction[J].Neural Computing and Applications, 2023, 35(18): 13109-13118.

[6] Challu C, Olivares K G, Oreshkin B N, et al. Nhits: Neural hierarchical interpolation for time series forecasting[C]//Proceedings of the AAAI conference on artificial intelligence. 2023, 37(6): 6989-6997.

[7] Guo C, Fan B, Zhang Q, et al. Augfpn: Improving multi-scale feature learning for object detection[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2020: 12595-12604.

[8] Van Der Donckt J, Van Der Donckt J, Van Hoecke S. tsdownsample: High-performance time series downsampling for scalable visualization[J]. SoftwareX, 2025, 29: 102045.

[9] Taud H, Mas J F. Multilayer perceptron (MLP)[M]//Geomatic approaches for modeling land change scenarios. Cham: Springer International Publishing, 2017: 451-455.

[10] Sun Y, Zhou Q, Sun L, et al. CNN–LSTM–AM:A power prediction model for offshore wind turbines[J]. Ocean Engineering, 2024, 301: 117598.

[11] Wu H, Hu T, Liu Y, et al. Timesnet: Temporal 2d-variation modeling for general time series analysis[J]. arXiv preprint arXiv:2210.02186, 2022.

[12] Wang J, Zhang L, Li X, et al. ULSeq-TA: Ultra-long sequence attention fusion transformer accelerator supporting grouped sparse softmax and dual-path sparse Layer Norm[J].IEEE Transactionson Computer Aided  Design of Integrated Circuits and Systems, 2023, 43(3): 892-905.

[13] Wu H, Xu J, Wang J, et al. Autoformer: Decomposition transformers with auto-correlation for longterm series forecasting[J].Advances in neural information processing systems, 2021, 34: 22419-22430.

[14] Zhou T, Ma Z, Wen Q, et al. Fedformer: Frequency enhanced decomposed transformer for long-term series forecasting[C]//International conference on machine learning. PMLR, 2022: 27268-27286.

[15] Yang M, Wang D, Zhang W. A short-term wind power prediction method based on dynamic and static feature fusion mining[J]. Energy, 2023, 280: 128226. 

[16] Kocak A, Erdem E, Erdem A. A gated fusion network for dynamic saliency prediction[J]. IEEE Transactions on Cognitive and Developmental Systems, 2021, 14(3): 995-1008.

[17] Wu H, Hu T, Liu Y, et al. Time-Series-Library: A comprehensive benchmark for deep time series forecasting[J/OL]. GitHub Repository, 2022.

[18] Covert I C, Qiu W, Lu M, et al. Learning to maximize mutual information for dynamic feature selection[C]//International Conference on Machine Learning. PMLR, 2023: 6424-6447.

[19] QianH,Wen Q,Sun L,et al.Robustscaler:Qos-aware autoscaling for complex workloads[C]//2022 IEEE 38th International Conference on Data Engineering (ICDE). IEEE, 2022: 2762-2775.

[20] Arefi M, Khammar A H. Nonlinear prediction of fuzzy regression model based on quantile loss function[J]. Soft Computing-A Fusion of Foundations, Methodologies and Applications, 2024, 28(6).