基于TWSVM和模糊的木质采暖地板蓄热温度预测模型

Temperature Prediction Model of Heat Storage of Wooden Flooring for Ground with Heating System Based on TWSVM and Fuzzy

【目的】围绕木质地板蓄热特性,针对木质地板蓄热后生成的温度场分布建立基于孪生支持向量机(TWSVM)和模糊算法的温度组合预测模型,为后续研究木质地板的蓄热规律提供有效分析手段。【方法】首先,将加热到设定温度的木质地板样本推送至地采暖地板蓄热性能检测仪器的检测腔内自由放热,利用仪器内呈多层环状分布的温度传感器阵列在时间和空间维度上动态提取温度值,并进行滤波降噪和归一化处理。其次,针对建模数据过多导致的TWSVM计算复杂度迅速膨胀问题,将试验数据均匀分块,每个分块数据中的验证集样本随机提取,剩余为训练集样本,采用TWSVM分别训练每个训练集样本,并用对应样本中的验证集进行泛化性验证,运用网格搜索法对TWSVM模型的核函数参数σ、惩罚参数γ和松弛因子ξ进行寻优。最后,基于模糊原理,对验证样本的输入空间构建高斯隶属度函数,并应用隶属度函数将模型预测结果进行模糊叠加,叠加后的输出作为模型最终训练结果。【结果】基于模糊的TWSVM方法预测时间维度下不同样本温度值的最大拟合度为99.59%,最小为98.92%,最长建模时间为186.90 s,最短建模时间为64.39 s;预测空间维度下不同样本温度值的最大拟合度为99.23%,最小为98.96%,最长建模时间为274.37 s,最短建模时间为93.30 s。【结论】TWSVM在计算中涉及矩阵求逆问题,适合对维数较小的数据样本进行建模,由于本研究木质地板蓄热特性需要的温度数据量较大,因此采用TWSVM直接对该试验数据进行建模具有较大局限性;引入模糊方法后,先将温度数据分别在时间和空间维度上分割成多个小的训练样本,然后对每个训练样本分别采用TWSVM建模和训练,根据模糊规则,以每个温度点在模糊函数上的隶属度叠加值来确定最终预测结果,可提高TWSVM方法建模的适应范围,并充分发挥其快速性和泛化性优势。

【Objective】 With regard to the heat storage characteristics of wooden flooring, a temperature prediction model based on twin support vector machine（TWSVM）and fuzzy algorithm was established for the distribution of temperature field of wooden flooring heat storage, which provides an effective analysis method for further research on heat storage of wooden flooring.【Method】 Firstly, the wood samples which are heated to the set temperature are pushed into the detection cavity of the heating floor storage performance testing instrument for free heat dissipation, and the temperature sensor array with multi-layer annular distribution characteristics in the instrument is used to dynamically record the temperature data on the time dimension and spatial dimension, then filtering noise reduction and normalization processing are carried out. Secondly, aiming at the problem that too much modeling data will lead to the rapid expansion of the computational complexity of the twin support vector machine, the test data is evenly partitioned in the paper, and the verification set sample are randomly extracted from each block data, and the remaining ones are training samples. Based on the different training samples, the TWSVM model is trained, and the generalization performance of the model is verified by the verification set in the corresponding samples. The kernel function parameters σ, penalty parameters γ and relaxation factors ξ of the model are optimized by the grid search method. Finally, based on the fuzzy principle, the Gaussian function is constructed for the input space of the test sample.The prediction result of all models are superimposed with the corresponding membership weight values to generate the final training result of the model.【Result】 The result show that when TWSVM and fuzzy method are used to predict the temperature values of different samples in time dimension, the maximum fitting degree is 99.59%, the minimum is 98.92%, the longest and shortest modeling time is 186.90 s and 64.39 s, respectively. When the temperature values in spatial dimension are predicted, the maximum fitting degree is 99.23%, the minimum is 98.96%, the longest and shortest modeling time is 274.37 s and 93.30 s, respectively.【Conclusion】 Because the TWSVM method involves matrix inversion in the computation process, it is only suitable for processing small data samples. In this experiment, because of the large amount of temperature data needed to study the storage characteristics of wooden flooring, there is a limitation to directly use TWSVM to model the experimental data. After introducing the fuzzy method , the temperature data are divided into several small training samples in time and spatial respectively, then each training sample is modeled and trained by TWSVM, respectively. According to fuzzy rules, the final prediction result is determined by the membership value of each temperature point on the fuzzy function. The above method can improve the adaptation range of TWSVM modeling, and give full play to its fast and generalization advantages.