中国是全球受台风灾害影响最严重的国家之一，而广西是中国受台风影响最严重的地区之一，年均受到5次台风影响，最高达9次，台风灾害给广西地区带来了巨大的损失，严重阻碍了广西社会经济的高质量发展。本研究选取2001-2020年影响广西并造成一定经济损失且有较完整记录的40个历史台风灾害事例，从致灾因子、承灾体、防灾减灾能力三个指标层出发，以12个台风影响因子作为神经网络模型的输入要素，使用4种神经网络结合三次样条插值法进行数据扩增构建了广西台风灾损智能预测模型，并以台风“彩虹”为例探索了该模型的实战应用潜力，实现广西台风灾损动态预测。结果表明，性能表现最好的为GA-BP(genetic algorithm - back propagation)神经网络模型，模型训练集R方为0.984 7、RMSE(root mean squared error)为0.457 3，测试集R方为0.960 3、RMSE为0.829 5，预测结果与实际台风灾损情况比较接近，证明了该模型的有效性。

China is one of the countries most severely affected by typhoon disasters globally, and Guangxi is one of the regions in China most severely affected by typhoons. On average, Guangxi is affected by typhoons five times a year, with a maximum of nine times. Typhoon disasters have brought huge losses to the Guangxi region and seriously hindered the high - quality development of Guangxi's social economy. In this study, 40 historical typhoon disaster cases that affected Guangxi, caused certain economic losses, and had relatively complete records from 2001 - 2020 were selected. Starting from the three index layers of hazard - causing factors, disaster - bearing bodies, and disaster prevention and mitigation capabilities, 12 typhoon - influencing factors were taken as the input elements of the neural network model. Four kinds of neural networks combined with the cubic spline interpolation method were used for data amplification to construct an intelligent prediction model for typhoon disaster losses in Guangxi. Taking Typhoon Mujigae as an example, the practical application potential of this model was explored to realize the dynamic prediction of typhoon disaster losses in Guangxi. The results show that the GA - BP(genetic algorithm - back propagation) neural network model has the best performance. The R - squared value of the model training set is 0.9847, the RMSE(root mean squared error) is 0.4573, the R - squared value of the test set is 0.9603, and the RMSE is 0.8295. The prediction results are relatively close to the actual typhoon disaster loss situation, which proves the effectiveness of this model.