Is this the most basic lstm model?

[qqzj1001]

Hello, I am a beginner. Is this the most basic lstm model? Can I use this as a baseline? Or should a more complex model be used?

[TudorOrban]

Hello there. This is indeed a pretty basic model. On whether you should use it, it depends on what you want to do. If you wish to learn about LSTM's, it might be useful, although I would personally check out the github topic, there are plenty of better models out there. If you are interested particularly in ML for climate predictions, you can try extending this model, perhaps by using the monthly means in the NASA dataset. More advanced models in climate science research use Convolutional LSTM's or other CNN's to take spatial data into account (see this for instance, though you'd be venturing into math-heavy territory).

[qqzj1001]

你好。这确实是一个相当基本的模型。关于你是否应该使用它，这取决于你想做什么。如果你想了解LSTM，它可能很有用，尽管我会亲自查看github主题，但有很多更好的模型。如果您对ML的气候预测特别感兴趣，您可以尝试扩展此模型，也许可以使用美国宇航局数据集中的月度平均值。气候科学研究中更先进的模型使用卷积LSTM或其他CNN来考虑空间数据（例如，请参阅此，尽管您将冒险进入数学重领域）。

Thanks for your reply and suggestions. My current research direction is indeed related to climate change. Maybe I will consider using more advanced models, but the basic model is relatively easy to understand for beginners. At present, I don’t know how to extend the model. All in all your code is very useful to me, thanks

[TudorOrban]

Great to hear you're pursuing this research direction. Regarding extending the model, I've used the Zonal Annual Means to train the model, which was easy to work with but it's too small to get good predictions, which is why you might see the model overfit on different training runs. Using the monthly means dataset from the link above should allow you to increase the model size and obtain better results. Another idea is to somehow "inform" the model of what the features in the annual dataset mean: geographical areas (northern hemisphere, 24N-90N etc.). This might require a convolutional approach. I'm not actively working on this project but if you will have any more questions, feel free to ask me.

[qqzj1001]

很高兴您听到正在追求这个研究方向。关于扩展模型，我使用了区域_年均值_来训练模型，这很容易使用，但它太小而无法获得良好的预测，这就是为什么您可能会看到模型在不同的训练运行中过度重复。使用上面链接中的_每月_均值数据集应该可以让您增加模型大小并获得更好的结果。另一个想法是采用某种方式“告知”模型年度数据集中的特征的意义：地理区域（北半球、北纬24度至北纬90度等）。这可能需要外部方法。我没有积极参与这个项目，但如果您还有其他问题，请随时问我。

Okay, thank you very much for your answer. I look forward to learning more from you.