Abstract In the present study, moisture content evolution of cylindrical quince slices during convective drying was modelled by using artificial neural networks (ANN). Quince slices with an average initial moisture content of 81% in wet basis (w.b.) or 4.27 kgwater/kgdry matter in dry basis (d.b.), were dried in a laboratory thermal convective dryer and experimental data of moisture content versus drying time was obtained for nine measurement groups of 40, 50 and 60 °C drying air temperature and 1, 2 and 3 m/s airflow velocity respectively. Different topologies of multilayer perceptron (MLP) ANN models containing a single or two hidden layers with a different number of hidden neurons and different types of transfer functions, have been investigated for predicting the moisture content evolution during drying. A group k-fold cross validation iteration procedure was performed for each developed ANN structure, in order to assess each model’s ability to estimate the moisture content of quinces on unseen data of air-drying temperature and airflow velocity combinations held out of the training process. For the cross validation of the developed ANN models, appropriate statistical evaluation indices were applied. The best performed ANN model based on the cross validation score metrics, contained two hidden layers with the sigmoid, softplus transfer functions and was composed by 90 artificial hidden neurons in each of the two hidden layers. A satisfying agreement of predictions with the experimental data was noticed, achieving coefficients of determination (R2) greater than 99% and root mean square error (RMSE) values less than 0.08 kgwater/kgdry matter.

中文翻译：

---

圆柱形木瓜切片对流干燥含水率演化的人工神经网络建模

摘要 在本研究中，通过使用人工神经网络 (ANN) 模拟了对流干燥过程中圆柱形木瓜切片的水分含量演变。将湿基 (wb) 的平均初始水分含量为 81% 或干基 (db) 的平均初始水分含量为 4.27 kg 水/kg 干物质 (db) 的木瓜切片在实验室热对流干燥机中干燥，并获得水分含量与干燥时间的实验数据9 个测量组，分别为 40、50 和 60 °C 的干燥空气温度和 1、2 和 3 m/s 的气流速度。多层感知器 (MLP) ANN 模型的不同拓扑结构包含一个或两个隐藏层，具有不同数量的隐藏神经元和不同类型的传递函数，已被研究用于预测干燥过程中的水分含量演变。对每个开发的 ANN 结构执行了一组 k 折交叉验证迭代程序，以评估每个模型在训练过程中保留的风干温度和气流速度组合的看不见的数据上估计木瓜水分含量的能力。对于开发的 ANN 模型的交叉验证，应用了适当的统计评估指标。基于交叉验证分数指标的最佳 ANN 模型包含两个带有 sigmoid 和 softplus 传递函数的隐藏层，并由两个隐藏层中的每一个中的 90 个人工隐藏神经元组成。注意到预测与实验数据的令人满意的一致性，实现了大于 99% 的决定系数 (R2) 和小于 0.08 千克水/千克干物质的均方根误差 (RMSE) 值。