|
|
Post by isaacmg on Apr 4, 2016 19:57:49 GMT -5
Hey everybody, so over the past year we (at PaddleSoft) have collected a massive amount data on the Kenduskeag and its flow patterns. We used the data to train a neural network which given a weather forecast (i.e. rainfall, temperature) and current flow can predict n hours in the future. We will have a tool up and running (online) which shows our realtime projections very soon but in the interim you can read our article which details some of our findings about how the Kenduskeag responds to "rainfall events." I'm also happy to answer any questions you might have. Link to articleURL in case that doesn't work. medium.com/@paddlesoft/forecasting-the-flow-of-the-kenduskeag-a731c5d35aa6#.egqa81dtm |
|
|
|
Post by isaacmg on Apr 9, 2016 22:58:51 GMT -5
Okay finally got my prediction system online. It is not automated (yet) so until then it'll be updated twice a day. Once I can finally deploy my MATLAB code to my web server it will be updated once per hour. On the test data the NN I trained to predict height was a bit more accurate than the one I used for cfs. Current predictions for the Kenduskeag race seem to be around 5.67 and 1250CFS. Anyways here are the links isaac26.com/kenduskeag.html isaac26.com/kenduskeag2.html |
|
|
|
Post by Curious on Apr 12, 2016 5:32:45 GMT -5
Okay finally got my prediction system online. It is not automated (yet) so until then it'll be updated twice a day. Once I can finally deploy my MATLAB code to my web server it will be updated once per hour. On the test data the NN I trained to predict height was a bit more accurate than the one I used for cfs. Current predictions for the Kenduskeag race seem to be around 5.67 and 1250CFS. Anyways here are the links isaac26.com/kenduskeag.html isaac26.com/kenduskeag2.htmlThe original prediction is way off from current prediction. Which is right? |
|
|
|
Post by cdalton on Apr 12, 2016 8:47:40 GMT -5
The precipitation forecast keeps getting revised downward, so probably the latest model run (229cfs as of now) reflects that.
Everybody do a rain dance. Today's rain is all we'll get. 229 is not crazy low like 2012, but it's lower than 2010, which was described as "a slog".
|
|
|
|
Post by isaacmg on Apr 12, 2016 15:41:39 GMT -5
The current prediction is the closest. The model pulls whatever is the latest hourly precipitation forecast from WunderGround and then feeds that into the NN. The reason there has been so much change in the projections is due to changes in the hourly precipitation forecast. As I stated in the article the model is only as good as the weather forecast. So as the weather forecast changes the model will change. In terms of the models accuracy I think that the height model is a bit more accurate than cfs. It was generally only off an average of .3ft on the test data when given the rainfall/temp. You can keep checking the website for the most recent projections. I'm trying to get the website to update hourly (as the forecast changes) however this is difficult due to the inability to deploy MATLAB code. Until then I will keep updating the website manually ever few hours.
|
|
|
|
Post by johnfrombangor on Apr 15, 2016 8:26:35 GMT -5
229? now how do you like them apples... hope the machine's wrong...
|
|
|
|
Post by charlie on Apr 15, 2016 11:18:47 GMT -5
Probably be right around 400 for tomorrow.
|
|
|
|
Post by isaacmg on Apr 15, 2016 12:45:20 GMT -5
Try refreshing the page it's currently predicting 319cfs however, that is probably on the low side. I would guess probably around 450cfs and 4.9ft. The training set I used unfortunately was all taken from the summer/fall when the level dropped faster due to warmer weather and it being dryer overall. I think this is causing it over predict the drop in CFS. Hopefully, some more training data and parameter tuning will fix this problem in the future...
|
|
