To encourage proper employee scheduling for managing crew load, restaurants need accurate sales forecasting. This paper proposes a case study on many machine learning (ML) models using real-world sales data from a mid-sized restaurant. Trendy recurrent neural network (RNN) models are included for direct comparison to many methods. To test the effects of trend and seasonality, we generate three different datasets to train our models with and to compare our results. To aid in forecasting, we engineer many features and demonstrate good methods to select an optimal sub-set of highly correlated features. We compare the models based on their performance for forecasting time steps of one-day and one-week over a curated test dataset. The best results seen in one-day forecasting come from linear models with a sMAPE of only 19.6%. Two RNN models, LSTM and TFT, and ensemble models also performed well with errors less than 20%. When forecasting one-week, non-RNN models performed poorly, giving results worse than 20% error. RNN models extended better with good sMAPE scores giving 19.5% in the best result. The RNN models performed worse overall on datasets with trend and seasonality removed, however many simpler ML models performed well when linearly separating each training instance.

Methodologies inspired by physics-informed neural networks (PINNs) were used to forecast observations recorded by stationary ocean buoys. We combined buoy observations with numerical models to train surrogate deep learning networks that performed better than with either data alone. Numerical model outputs were collected from two sources for training and regularization: the hybrid circulation ocean model and the fifth European Centre for Medium-Range Weather Forecasts (ECMWF) reanalysis experiment. A hyperparameter determines the ratio of observational and modeled data to be used in the training procedure, so we conducted a grid search to find the most performant ratio. Overall, the technique improved the general forecast performance compared with nonregularized models. Under specific circumstances, the regularization mechanism enabled the PINN models to be more accurate than the numerical models. This demonstrates the utility of combining various climate models and sensor observations to improve surrogate modeling.

Methodologies inspired by physics-informed neural networks (PINNs) were used to forecast observations recorded by stationary ocean buoys. We combined buoy observations with numerical models to train surrogate deep learning networks that performed better than with either data alone. Numerical model outputs were collected from two sources for training and regularization: the hybrid circulation ocean model and the fifth ECMWF reanalysis experiment. A hyperparameter determines the ratio of observational and modeled data to be used in the training procedure, so we conducted a grid search to find the most performant ratio. Overall, the technique improved the general forecast performance compared with nonregularized models. Under specific circumstances, the regularization mechanism enabled the PINN models to be more accurate than the numerical models. This demonstrates the utility of combining various climate models and sensor observations to improve surrogate modeling.

When forecasting fixed-location observation nodes with statistical surrogate models, combining datasets during training loss calculation has been shown to improve model accuracy. However, traditional methods for tuning the data ratio, such as grid search or random search, are computationally prohibitive. An alternative online methodology for optimizing the data during training has been previously investigated. While both approaches have been independently validated, they have never been directly compared to each other or to other search techniques. This paper presents a direct comparison to evaluate whether the online approach can serve as a viable replacement for conventional search methods. The Cahn-Hilliard physical equation provides a controlled testing environment for this analysis. The results show that the optimization algorithm may require additional improvements before an out-of-the-box approach is appropriate. However, using the derived optimal hyperparameter in an offline setup provides an improvement in accuracy, which implies the methodology is worthwhile when under time constraints.

Container salvaging is becoming common in maritime industry, and we are more interested in recycling containers from the sea. Often, containers are often loaded with either valuable or harmful substances and must be retrieved. Therefore, there is motivation for model tests and numerical calculations on container models falling into water. 3D-printed container models, representing different loading conditions, have been dropped into the towing tank at the University of New Orleans (UNO). Videos of these dropped containers are analyzed to classify the initial angle of the container dropped into water. The containers were dropped 57 times each at 0°, 45°, and 90° angles into the tank. Preprocessing tasks are conducted on the videos to prepare for model training. The two-phase angle classification approach uses a pre-trained ResNet50 model, trained on ImageNet, as a feature extractor to produce latent features to train a gated recurrent unit (GRU) model for classifying the entry angle of the dropped object. Initial results have shown it is possible for a model to classify the angles of dropped objects into water with up to 95% accuracy. Furthermore, the data obtained from these small-scale experiments will be integrated into the existing in-house tool “DROBS”.