基于小样本迁移学习的跨地点菇房温度预测模型研究

Cross-location Mushroom House Temperature Prediction Model Based on Small Sample Transfer Learning

摘要:

菇房温度精准预测是保证食用菌工厂化高效生产的关键，现有预测模型在面向不同地点菇房实际应用中缺乏普适性。本文考虑菇房设备对环境变化的扰动影响，融合设备运行状态特征建立一种基于时序卷积网络和长短期记忆网络（Temporal convolution network-long short-term memory，TCN-LSTM）的菇房温度时序预测模型，采用TCN在时间维度上提取特征局部信息，利用LSTM捉时序数据的长期依赖关系。与未融合设备工作状态特征的模型相比，当预测时长为1、2、3h时，TCN-LSTM模型MSE分别降低17.1%、35.7%、44.1%，MAE分别降低4.3%、28.0%、38.0%；结果表明考虑融合设备工作状态提升了模型预测性能。与其他不同浅层和深度学习模型相比，TCN-LSTM模型在不同预测时长下均具有最佳的预测精度，当预测时长为小于，3h模型R2不小于0.982、MSE和MAE不大于0.57℃，可满足菇房环境调控对温度预测精度和时长要求。基于迁移学习的预训练及微调方式在小样本数据集中调整网络参数，实现面向不同地点的菇房温度预测模型快速构建。结果表明，当预测时长为1、2、3h时，以不同地点数据作为目标域建立的预测模型在测试集中R2不小于0.912，MSE不大于4.02℃，MAE不大于2.01℃，因此，在小样本条件下基于迁移学习构建的面向不同地点模型可实现不同步长下温度精准预测。

Abstract:

Accurate temperature prediction in mushroom houses is crucial to ensuring the efficient industrial production of edible mushrooms. However, existing predictive models often lack generalizability when applied to mushroom houses located in different regions. Taking into account the disruptive effects of environmental changes caused by equipment in mushroom cultivation houses, and equipment operation state features was integrated to develop a temperature prediction model based on the temporal convolution network and long short-term memory model (TCN-LSTM). This model used TCN to extract local information along the temporal dimension, while LSTM captured the long-term dependencies of time series data. Compared with models that did not integrate device operating state features, the TCN-LSTM model that incorporated these features reduced the MSE by 17.1%, 35.7%, and 44.1%, and reduced MAE by 4.3%, 28.0%, and 38.0% for prediction horizons of 1hour, 2hours, and 3hours, respectively. This result indicated that incorporating equipment operating states had a significantly positive effect on the prediction performance. Compared with other shallow and deep learning models, the TCN-LSTM model achieved the best prediction accuracy for different prediction horizons, with R2 no less than 0.982, and both MSE and MAE no more than 0.57℃ for horizons up to 3 hours, satisfying the requirements of accuracy and duration in temperature predictions for mushroom room environmental control. This study employed transfer learning via pre-training and fine-tuning to adjust network parameters in small sample datasets, achieving rapid construction of temperature prediction models for mushroom rooms at different locations. The results indicated that for prediction horizons of 1hour, 2hours, and 3hours, the prediction models built using different locations as target domains achieved R2 values no less than 0.912, MSE values no more than 4.02℃, and MAE values no more than 2.01℃ in the test set. These results suggested that under small sample conditions, the temperature models constructed for different locations using transfer learning can achieve accurate temperature predictions at different steps.