-
Notifications
You must be signed in to change notification settings - Fork 0
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
comitting readme of submission for enefit competition
- Loading branch information
1 parent
40ae816
commit 96cda0d
Showing
1 changed file
with
4 additions
and
0 deletions.
There are no files selected for viewing
4 changes: 4 additions & 0 deletions
4
Kaggle_Enefit_Multi-Step_Forecast_DNN_MultivariateTimeSeries/README.md
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,4 @@ | ||
### Enefit - Predict Energy Behavior of Prosumers Competition | ||
This folder contains notebooks I submitted to the [Enefit - Predict Energy Behavior of Prosumers competition](https://www.kaggle.com/competitions/predict-energy-behavior-of-prosumers). Instead of relying on a public baseline notebook, I developed my own preprocessing pipeline and experimented with various deep neural network (DNN) structures. I designed a DNN with encoder blocks that provides decent multistep predictions. This notebook includes the complete pipeline and the actual model, capable of producing predictions via a simple API call. My attempts to fine-tune the model were limited to adjusting the number of encoder blocks, as last-minute tuning of dropout rates did not yield improvements. | ||
|
||
The gather layers in the functional model might seem unconventional. Initially, I didn't plan to partition the input, but poor performance with a single block led me to distribute and partition the data. Although using hardcoded indices was inelegant, it prevented crashes during the load_model phase. A multi-input model would have been more elegant, avoiding the need for gather layers and input division during inference. Additionally, using attention on numeric and time features was suboptimal; attention should have been reserved for categorical features. At the time, I lacked the knowledge to implement multi-input models or build deep networks with model subclassing. This project motivated me to study advanced techniques in TensorFlow, where I learned how to implement these approaches. Although my model didn't excel on the leaderboard, the experience was invaluable for my learning and development. This notebook is publicly available and open for comments on [Kaggle](https://www.kaggle.com/code/pelinkeskin/multivariate-multistep-dnn-cnn-mah-lstm). |