Article: Algorithm application and verification for sports training load feature analysis Journal: International Journal of Information and Communication Technology (IJICT) 2026 Vol.27 No.30 pp.73 - 92 Abstract: This study intends to build a hybrid deep learning model based on transformer and bidirectional long short-term memory network (BilSTM) using high-precision multi-source data of NBA players. First, this study identifies critical load periods and captures sequence global dependence using transformer encoder's multi-head self-attention (MSA). Then, the BiLSTM layer is used to enhance the bidirectional time-series context modelling, and finally the scenario classification results are output. The experimental results demonstrate that in the load trend prediction task, the model's mean squared error (MSE) decreases by 23.3% compared to traditional long short-term memory (LSTM), and by 13.8% compared to the pure transformer model. While the R<SUP align="right"><SMALL>2</SMALL></SUP> increases to 0.944, with an improvement of approximately 1.7%, indicating superior fitting accuracy and stability. This model demonstrates significant advantages in both recognition accuracy and generalisation capability. Overall, the proposed method is more sensitive and stable in capturing exercise load's complex temporal dependencies. Inderscience Publishers - linking academia, business and industry through research

Title: Algorithm application and verification for sports training load feature analysis

Authors: Jingjing Duan

Addresses: School of Hebei Software Institute, Urban, 071000, China

Abstract: This study intends to build a hybrid deep learning model based on transformer and bidirectional long short-term memory network (BilSTM) using high-precision multi-source data of NBA players. First, this study identifies critical load periods and captures sequence global dependence using transformer encoder's multi-head self-attention (MSA). Then, the BiLSTM layer is used to enhance the bidirectional time-series context modelling, and finally the scenario classification results are output. The experimental results demonstrate that in the load trend prediction task, the model's mean squared error (MSE) decreases by 23.3% compared to traditional long short-term memory (LSTM), and by 13.8% compared to the pure transformer model. While the R² increases to 0.944, with an improvement of approximately 1.7%, indicating superior fitting accuracy and stability. This model demonstrates significant advantages in both recognition accuracy and generalisation capability. Overall, the proposed method is more sensitive and stable in capturing exercise load's complex temporal dependencies.

Keywords: training load analysis; transformer model; bidirectional long short-term memory network; time series data mining; competitive sports.

DOI: 10.1504/IJICT.2026.152657

International Journal of Information and Communication Technology, 2026 Vol.27 No.30, pp.73 - 92

Received: 14 Oct 2025
Accepted: 11 Nov 2025
Published online: 01 Apr 2026 *

Title: Algorithm application and verification for sports training load feature analysis

Keep up-to-date