vast fable
I am working on a time series forecasting project where I aim to predict load power consumption based on various features including date and time, holiday indicators, and seasonal information. Despite preprocessing steps such as outlier detection, feature engineering, and trying different machine learning algorithms including LSTM,CNN, ConvLSTM models, I am struggling to achieve satisfactory performance. I have implemented outlier detection techniques to identify and remove outliers from the dataset, particularly in the target column. However, even after outlier removal, the model performance remains suboptimal.I have also performed feature engineering, by extracting date and time features from the datetime column and encoded categorical features such as seasons. These features were expected to improve model accuracy, but the results have been disappointing.I have trained models with different architectures and hyperparameters, but I have not seen significant improvements in performance. I have tried tuning learning rates, adjusting layer sizes, and regularization techniques, but the model still fails to capture the underlying patterns effectively.I have experimented with data transformations such as log transformation and square root transformation to handle non-linear relationships and improve the distribution of the target variable which is right skeewed as it clear in the picture. However, these transformations did not lead to significant enhancements in model performance. you find attached the distribution of data and differents processing steps applied.
orchid saffron
