This paper illustrates how Machine Learning techniques can be used to assess the impact of environmental protocols that are sparsely activated over time. A case study is analysed: the impact of a protocol that sets traffic restrictions on NO₂ levels in Madrid’s urban area. The protocol specifies that these restrictions are active only when NO₂ level reaches above a certain threshold. Since the protocol was first enacted, in 2017, restrictions have only been active for 59 days, never longer than ten consecutive days. Cross effects are identified: the protocol magnifies the effect of other relevant features, especially wind speed. Pollution decreases with wind speed. Consequently, episodes of high pollution levels generally occur when wind speed is low. The analysis shows that precisely at these low values of wind speed, an increase in the wind speed has a higher effect on pollution when the protocol is activated than when it is not. The analysis is carried out at a measuring station level, considering eight representative stations. The referred cross effect is clearer at centrally located stations across Madrid. Cross effects of the protocol with other weather features are weaker than that with wind speed. The assessment is based on the computation of Shapley values for classification trees that are built using XGBoost and rolling windows.

本文说明了机器学习技术如何用于评估随时间推移稀疏激活的环境协议的影响。案例研究分析：设置无流量限制协议的影响，_2个水平在马德里市区。该协议规定只有在NO ₂时这些限制才有效级别达到某个阈值以上。自该协议于2017年首次颁布以来，限制措施仅生效了59天，从未超过连续十天。识别出交叉影响：该协议会放大其他相关功能的影响，尤其是风速。污染随风速降低。因此，当风速低时，通常会发生高污染水平。分析表明，正是在这些低风速值下，激活该协议时，风速的增加对污染的影响要比未激活时更高。考虑到八个代表站，该分析是在测量站级别进行的。在马德里市中心的车站，这种交叉效应更为明显。该协议与其他天气特征的交叉影响要弱于风速。该评估基于使用XGBoost和滚动窗口构建的分类树的Shapley值的计算。