是否有任何损失函数更适合于具有多个零的数据？

Question

我一直在使用多层LSTM模型来预测汽车充电站每小时的能耗。由于数据的性质，在使用峰值之间有许多零。有没有办法使模型对非零值更敏感？我一直在使用标准损失函数来提高精度，但我觉得非零预测的实际误差似乎被预测零的“精度”所掩盖

作为参考，我使用了标准的RMSE损失函数和标准的TensorFlow LSTM模型

数据集的外观：

17,2017-09-01 17:00:00,0.0
18,2017-09-01 18:00:00,0.0
19,2017-09-01 19:00:00,0.0
20,2017-09-01 20:00:00,15.13000000000007
21,2017-09-01 21:00:00,0.0
22,2017-09-01 22:00:00,0.0
23,2017-09-01 23:00:00,0.0
24,2017-09-06 00:00:00,0.0
25,2017-09-06 01:00:00,0.0
26,2017-09-06 02:00:00,0.0
27,2017-09-06 03:00:00,0.0
28,2017-09-06 04:00:00,0.0
29,2017-09-06 05:00:00,0.0
30,2017-09-06 06:00:00,0.0
31,2017-09-06 07:00:00,0.0
32,2017-09-06 08:00:00,0.0
33,2017-09-06 09:00:00,0.0
34,2017-09-06 10:00:00,0.0
35,2017-09-06 11:00:00,0.0
36,2017-09-06 12:00:00,10.959999999999999
37,2017-09-06 13:00:00,0.0
38,2017-09-06 14:00:00,0.0

我感谢任何关于这个问题的指导