I have weather data for the 40 previous years and for each year I have information about the hurricane season (intensity, number of active days, casualties,...). My ultimate goal would be to forecast this season's hurricane intensity,... from 2019 weather data.
More formally, for each year $i$, I have $x^{(i)}_1, x^{(i)}_2, ..., x^{(i)}_T$ a multivariate time-series ($x_j^{(i)} \in {R}^{k\times d}$ and it contains $k$ variables (temperature, sea level pressure,...) at $d$ different points (given by longitude and latitude) at the earth's surface at year $i$ and hour $j$ so that $T = 8760$ if we consider all observations for all hours in one year) and I have a (known) vector $y^{(i)}$ containing the hurricane information for that year so that the data from the previous years is "labled" and that should be the basis for the learning process.
The overall dimension here is huge. Is there a way to reduce the dimension (temporally, spatially and predictor variables-wise) and have an efficient algorithm that can forecast this year's vector $y^{(2019)}$ from data $x^{(2019)}_1, x^{(2019)}_2, ..., x^{(2019)}_T$ for each $T$: meaning that we want to be able to forecast $y$ with data up to april, up to may, up to june...with (intuitively) increasing accuracy.
I am familiar with techniques usually used in big dimension problems (PCA and Neural Networks mainly) but not when data is (space and time)-dependent.
Any help would be appreciated. Thanks.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.