This code is the repository for implementing algorithms for real-time prediction of macroscopic congestion from network state variables using Deep Learning. The corresponding work is detailed in the following papers:

• Sudatta Mohanty, Michael Cassidy, Alexey Pozdnukhov, Real-Time Macroscopic Congestion Prediction Using Deep Learning, Transportation Research Part c, 2017 (in preparation)
• Sudatta Mohanty, Alexey Pozdnukhov, GCN-LSTM Framework For Real-Time Macroscopic Congestion Prediction, Bay Area Machine Learning Symposium (BayLearn), 2017 (accepted)

There is implementation of Graph Convolutional Network (GCN) based on the opensource codebase corresponding to:

• Michaël Defferrard, Xavier Bresson, Pierre Vandergheynst, Convolutional Neural Networks on Graphs with Fast Localized Spectral Filtering, Neural Information Processing Systems (NIPS), 2016.

The Neural Attention Framework used is based on the concepts discussed in:

• Vasili Ramanishka, Abir Das, Jianming Zhang, and Kate Saenko, Top-down Visual Saliency Guided by Captions, Conference on Computer Vision and Pattern Recognition (CVPR), 2017

The base default test network is represented by the following figure:

To generate plots for test scenarios on test network:

• Run /src/test/python/model_test/test_run_scenarios.py
• Visualize output files in /src/test/python/model_test/scenario_<Scenario_Index>

.json config parameters:
{
"freeway_links_jam_density": List of jam density values (veh/km) for freeway links (between zones),
"check_queue_spillover": Update curves after checking queue spillover condition (true/false),
"plot_congestion_io_curves": Plot I-0 curves for congestion links (within zones) (true/false), "freeway_links_length": List of lengths (km) for freeway links(between zones),
"get_curves_data": Output .csv files for data generated for each curve during simulation (true/false),
"plot_demand_congestion_curves": Plot curves for demand vs time and congestion vs time (true/false),
"congestion_links_fftt": List of free flow travel times (mins) for congestion links (within zones),
"num_bins": Number of time bins within a day,
"freeway_links_fftt": List of free flow travel times (mins) for freeway links (between zones),
"demand_start_times": List of start times (minutes from midnight) for OD demand from each input zone,
"congestion_links_length": List of lengths (km) for congestion links (within zones),
"threshold_output_for_congestion": List of minimum congestion metric value (mins) below which congestion is treated as zero,
"min_intervals": Time interval (mins) for each time bin,
"congestion_links_capacity": List of capacity values (veh/hr) for congestion links (within zones),
"demand_slopes": List of slopes (veh/hr/min) of OD demand curves (equal positive and negative slopes),
"demand_end_times":List of end times (minutes from midnight) for OD demand from each input zone,
"file_directory": File path for base directory,
"num_zones": Total number of zones,
"plot_route_choice_io_curves": Plot IO curves for freeway links to compare route choice (true/false),
"check_route_choice": Update curves if users choose route to satisfy Deterministic User Equilibrium (DUE) (true/false),
"congestion_links_jam_density": List of jam density values (veh/km) for congestion links (within zones),
"freeway_links_capacity": List of capacity values (veh/hr) for freeway links (between zones),
"congestion_nn_smoothening_number": Number of nearest neighbors used for smoothening congestion vs time curves
}

Run /src/main/python/model_test/generate_model_data.py with following parameters:

• Config .json file path (For example, ./scenario_<Scenario_Index>/config_generate_model_data.json)
• Output .csv file path inside base directory defined in config .json file (For example lstm_input_data/case_<Case_Index>.csv)
• Variable slope across days - Boolean parameter which specifies whether there is 10% stdev in the slope of demand function across days or not (T/F)
• Variable start time across days - Boolean parameter which specifies whether there is 30 mins stdev in start time of demand function across days or not (T/F)
• Realistic demand - Boolean parameter which specifies start times from the three zones at 6AM +/- 30 mins, 7AM +/- 30 mins and 8AM +/ 30 mins respectively with 30 mins stdev in start time and 10% stdev in slope across days, if set T, then the previous two parameters are overwritten (T/F)
• Start day index
• End day index

The simplified Bay Area freeway network is represented by the following figure:

To generate data on simplified Bay Area freeway network for fitting models:
Contact corresponding author at sudatta.mohanty@berkeley.edu for H-W OD demand files, link count files and arrival count files generated for each scenario and run /src/main/python/bay_area_simplified_freeway_network/generate_model_data.py with following parameters:

• Config .json file path (For example, config_generate_model_data.json)

.json config parameters:
{
"num_zones": Total number of zones,
"num_links": Total number of (forward + backward) links,
"congestion_zone_nos": List of indices of zones for which congestion metric is calculated over time, "congestion_nn_smoothening_number": Number of nearest neighbors used for smoothening congestion vs time curves,
"threshold_output_for_congestion": List of minimum congestion metric value (mins) below which congestion is treated as zero,
"start_day": Start day index for simulation,
"end_day": End day index for simulation,
"num_profiles": Number of possible simulation scanerios for each day (currently 10),
"arrival_count_dir_base_path": Base file path for directory containing arrival counts .csv files,
"arrival_count_file_name": File name (with extension) for file containing arrival counts,
"link_count_dir_base_path": Base file path for directory containing link counts .csv files,
"link_count_file_name": File name (with extension) for file containing link counts,
"od_count_dir_base_path": Base file path for directory containing OD counts .csv files,
"od_count_dir_name": Directory name for OD counts .csv files,
"od_count_base_file_name": Base file name (to be appended with index) (without extension) for OD counts .csv files,
"output_file_name":Output file name (with .csv extension),
"min_intervals": Length (mins) of each time bin
}

2. Running 1-NN Model

Run /src/main/python/model_test/1NN.py with following parameters:

• .csv file path for file generated in the previous step

Output includes:

• display of plot showing comparison for congestion values for actual data and 1NN prediction
• display of plot showing RMSE vs iteration number
• average RMSE value

3. Running LSTM-only Model

Run /src/main/python/lstm.py with following parameters:

• Config .json file path (For example, ./model_test/config_lstm.json or ./bay_area_simplified_freeway_network/config_lstm_zone_0.json)

.json config parameters:
{
"input_file_path": File path of file generated in Step 1,
"input_data_column_index_ranges": List of numbers of even size for input column indices(Each consecutive pair is considered a start and end column for inputs into the model. For example [1,3,5,7] implies that the input columns are 1,2,3,5,6,7),
"output_column_index": Column index for target output variable,
"n_days": Total number of days of data for running model,
"learning_rate": Learning rate for gradient descent during optimization,
"batch_size": Batch size (continuous indices in time),
"dropout": Dropout rate (currently not implemented),
"n_input": Number of inputs (must be equal to number of input columns * 2),
"n_steps": Number of time intervals per input data,
"n_hidden": Number of hidden nodes per hidden layer,
"n_outputs": Number of time intervals per output data,
"min_lag": Number of time intervals between between first time bin of input and first time bin of output,
"n_layers": Number of hidden layers,
"display_step": Display outputs at this iteration interval,
"n_plot_loss_iter": If predicting less than a day, the loss is calculated only at the time intervals at n_steps * [n_plot_loss_iter,n_plot_loss_iter+1] (this is to ensure we compare apples to apples!),
"attention_display_step":Number of iterations after which attention is displayed (for all consecutive iterations until one full day is covered)
}

Output includes:

• display of plot showing comparison for congestion values for actual data and LSTM prediction every display_step (defined in config .json file) iterations
• display of plot showing RMSE vs iteration number (if prediction done for less than 1 day, then RMSE is calculated for part of day at n_steps * [n_plot_loss_iter, n_plot_loss_iter + 1] time bins) every display_step iterations
• average RMSE value every display_step iterations
• display of temporal attention model heatmap at attention_display_step iterations (the plots are displayed for all subsequent iterations until temporal attention for all times in the day are covered)
• display of spatial attention model heatmap at attention_display_step iterations (the plots are displayed for all subsequent iterations until spatial attention for all times in the day are covered)

4. Running GCN-LSTM Model

Run /src/main/python/graph_cnn_lstm.py with following parameters:

• Config .json file path (For example, ./model_test/config_graph_cnn_lstm.json or ./bay_area_simplified_freeway_network/config_graph_cnn_lstm_zone_0.json)

.json config parameters:
{
"input_file_path": File path of file generated in Step 1,
"shortest_path_adjacency_graph_file_path": File path of graph adjacency/weight matrix using shortest path to determine weights,
"trajectory_clustering_adjacency_graph_file_path": File path of graph adjacency/weight matrix using trajectory to determine weights,
"input_data_column_index_ranges": List of numbers of even size for input column indices(Each consecutive pair is considered a start and end column for inputs into the model. For example [1,3,5,7] implies that the input columns are 1,2,3,5,6,7),
"output_column_index": Column index for target output variable,
"n_days": Total number of days of data for running model,
"learning_rate": Learning rate for gradient descent during optimization,
"decay_rate": Decay rate of learning rate per iteration,
"momentum": Momentum value for learning rate of previous iteration,
"batch_size": Batch size (continuous indices in time),
"eval_frequency": Display outputs at this iteration interval,
"regularization": L2 regularizations of weights and biases,
"dropout": Dropout rate (currently not implemented),
"lstm_n_hidden": Number of hidden nodes per hidden LSTM layer,
"lstm_n_outputs": Number of time intervals per output data,
"lstm_min_lag": Number of time intervals between between first time bin of input and first time bin of output,
"lstm_n_layers": Number of hidden LSTM layers,
"display_step": Display outputs at this iteration interval,
"n_plot_loss_iter": If predicting less than a day, the loss is calculated only at the time intervals at n_steps * [n_plot_loss_iter,n_plot_loss_iter+1] (this is to ensure we compare apples to apples!),
"cnn_filter": Filter type for GCN (Currently only implemented "chebyshev5"),
"cnn_brelu": Bias and Relu for GCN (Currently only implemented "b1relu"),
"cnn_pool": Pooling for GCN (Currently only implemented maxpooling "mpool1"),
"cnn_num_conv_filters": List of number of convolutional filters for each layer of GCN,
"cnn_poly_order": List of polynomial orders (filter sizes) for each layer of GCN,
"cnn_pool_size": List of pooling size (1 for no pooling and power of 2 to make graph coarser),
"cnn_output_dim": Number of features per sample for GCN,
"attention_display_step":Number of iterations after which attention is displayed (for all consecutive iterations),
"graph_type": Type of graph used (knn / shortest_path / trajectory_clustering)
}

Output includes:

• display of relative location of points in the original graph
• display of spectrum of a Laplacians for original and coarsened graphs
• display of plot showing comparison for congestion values for actual data and LSTM prediction every display_step (defined in config .json file) iterations
• display of plot showing RMSE vs iteration number (if prediction done for less than 1 day, then RMSE is calculated for part of day at n_steps * [n_plot_loss_iter, n_plot_loss_iter + 1] time bins) every display_step iterations
• average RMSE value every display_step iterations
• display of temporal attention model heatmap at attention_display_step iterations (the plots are displayed for all subsequent iterations)
• display of spatial attention model heatmap at attention_display_step iterations (the plots are displayed for all subsequent iterations)