基于xLSTM-TA的稻谷多段逆流干燥预测模型

Prediction Model for Paddy Multi-stage Counter-flow Drying Based on xLSTM-TA

摘要:

稻谷多段逆流干燥工艺呈现长时序、多阶段多参数耦合以及工况突变等复杂特性,致使干燥状态的精准预测仍面临挑战,稻谷爆腰问题突出。本研究提出了一种基于时序注意力增强型扩展长短期记忆网络的稻谷多段逆流干燥预测模型（xLSTM-TA-PD）,其通过交替集成扩展长短期记忆网络（Extended long short-term memory,xLSTM）残差块构建网络主干,以提取干燥时序信息,增强模型的历史参数信息记忆能力,以及在应对参数突变时做出准确响应的能力。同时,引入时序注意力（Temporal attention,TA）机制改进网络结构,解决了网络对干燥阶段权重均一化的问题。以5HNH-15型连续式稻谷干燥机为研究对象开展模型验证试验,发现xLSTM-TA模型均方根误差为0.068%,决定系数为0.988,预测性能优于深床干燥解析模型及其他数据驱动模型。在阶跃工况下,模型最大预测误差为0.44%,仅2.7%的样本绝对误差超过0.2%,证明模型能够准确识别参数阶跃变化与干燥状态响应间的滞后特性。此外,模型平均推理时间为0.0566 s,速度较深床干燥解析模型提升7.7倍,满足实时预测需求。通过时序注意力权重可视化分析发现,模型对不同时间步的权重分配与实际干燥机理高度一致,显著增强了模型的可解释性。本研究为粮食干燥的智能化调控提供了高精度、低延迟的决策支持工具。

Abstract:

The multistage counter-flow drying process of paddy exhibits complex characteristics such as long-term temporal sequences,multi-stage and multi-parameter coupling,and sudden condition changes,which pose challenges for accurate prediction of drying states and exacerbate the issue of rice fissuring. A prediction model for multistage counter-flow of paddy was proposed based on a temporal attention-enhanced extended long short-term memory network (xLSTM-TA-PD). The model constructed its backbone by alternately integrating residual blocks of the extended long short-term memory (xLSTM) network,enabling effective extraction of temporal drying information,enhancing the model's ability to memorize historical parameter data,and improving its responsiveness to abrupt parameter changes. Furthermore,a temporal attention (TA) mechanism was incorporated to refine the network architecture,addressing the issue of uniform weighting across different drying stages. Validation experiments were conducted by using a 5HNH-15 continuous paddy dryer. The results showed that the xLSTM-TA model achieved a root mean square error of 0.068% and a coefficient of determination of 0.988,outperforming the deep-bed drying analytical model and other data-driven models. Under step-change conditions,the model's maximum prediction error was 0.44%,with only 2.7% of samples exhibiting an absolute error exceeding 0.2%,demonstrating its capability to accurately capture the lag characteristics between parameter step changes and drying state responses. Additionally,the model's average inference time was only 0.0566 s,representing a 7.7-fold speed improvement over traditional deep-bed drying analytical models,thus meeting the requirements for real-time prediction. Visualization analysis of temporal attention weights revealed that the model's weight distribution across different time steps aligned closely with the actual drying mechanism,significantly enhancing model interpretability. The research result can provide a high-precision,low-latency decision-support tool for intelligent control in grain drying processes.