目前，基于数据驱动的传统非线性系统建模方法主要着眼于模型拟合和应用，鉴于此，该文针对来自系统的某个重要参数受不确定性影响的最大容忍极限输出，构建基于最小二乘支持向量回归（LSSVR）的上边界模型，深入剖析了上边界模型的精度与稀疏特性之间的平衡关系对上边界模型输出的影响。首先，借助LSSVR的优化问题，将原等式线性约束变成满足上边界模型的不等式约束；接着，为提高模型精度，引入基于上边界模型预测输出与实际输出之间逼近误差的不等式约束；与此同时，借助LSSVR的权值二范数来控制上边界模型结构的复杂度，从而构建出新的目标函数，并与满足上边界模型的不等式约束建立新的优化问题；最后，对所建立的优化问题引入拉格朗日函数并借助Karush-Kuhn-Tucker最优化条件来获取相应的对偶优化问题，并将其转化为标准的二次规划问题来求解上边界模型的参数。由于所构造的新优化问题满足凸性，因此模型系数解是全局最优的。该文还通过实验分析了反映模型精度的最大逼近误差、均方根误差及反映模型稀疏特性的指标，论证了所提方法的有效性和优越性。

At present, traditional data-driven nonlinear system modeling methods primarily focus on model fitting and application. In this context, this paper constructs an upper boundary model based on least squares support vector regression (LSSVR) for the maximum tolerable output of a critical parameter from the system, which is influenced by uncertainty. The study delves into the relationship between the balance of model accuracy and sparsity, and its effect on the model output. First, by utilizing the optimization problem of LSSVR, the original equality linear constraints are transformed into inequality constraints that satisfy the upper boundary model. Next, to improve the model’s accuracy, an inequality constraint based on the approximation error between the predicted output of the upper bound model and the actual output is introduced. Meanwhile, the LSSVR’s weight L₂-norm is employed to control the complexity of the upper boundary model’s structure, thereby constructing a new objective function and establishing a new optimization problem that satisfies the inequality constraints of the upper bound model. Finally, the Lagrangian function is introduced into the optimization problem, and the Karush-Kuhn-Tucker conditions are used to derive the corresponding dual optimization problem, which is then converted into a standard quadratic programming problem to solve for the parameters of the upper bound model. Since the new optimization problem satisfies convexity, the solution for the model coefficients is globally optimal. The effectiveness and superiority of the proposed method are validated through experimental analysis, where the maximum approximation error, root mean square error, and sparsity-related metrics are used to reflect the model’s accuracy and sparsity characteristics.