Research on the Prediction Model of Vocational Education Learning Effect Based on TCN-LSTM

Abstract

In the digital transformation of vocational education, learning effect prediction is a key means to optimize teaching strategies. Aiming at the temporal series, nonlinearity and multimodal characteristics of learning behavior data, this study proposes a TCN-LSTM hybrid model that combines TCN expansion causal convolution and LSTM gated memory mechanism to simultaneously capture the local dependence and long-range correlation of learning behavior sequences, and solve the shortcomings of traditional methods. Data preprocessing uses sliding window, standardization and missing value interpolation, and constructs 1,205,600 valid samples based on 12 types of time series features (online logs, training records, etc.) of 12,580 learners. The model is trained with 10-fold cross-validation (training/test set 8:2), with Adam optimizer, 64/32 hidden layer nodes, 32 batch sizes, 50 iterations and 0.1 dropout rate, and the local features of TCN and LSTM global timing dynamics are fused through the attention mechanism. The results show that the comprehensive accuracy of TCN-LSTM is 93.2% (6.8%/4.1% higher than that of LSTM/TCN), MSE is 0.154 (55% lower than SVM), RMSE is 0.0632, MAE is 0.0427, R² is 0.9298, and the prediction accuracy of 1/3/5 week lag is 94.7%/87.4%/79.8% (The 3-week lag error is 32.6% lower than that of ARIMA), and the interdisciplinary prediction accuracy of mechanical/IT majors is 91.5%/89.8% (standard deviation 1.7%), which is better than FCN-LSTM with more parameters and slower speed. This study provides a new path for modeling learning behavior in vocational education, and verifies the effectiveness of hybrid neural networks in processing complex educational time series data.