Development of an Intelligent Freight Forecasting Model for Optimized Vessel Chartering and Bulk Cargo Procurement from overseas to East Coast of India

1. Executive Snapshot

Steel Authority of India Limited (SAIL) imports over 15–20 million metric tonnes of metallurgical coking coal annually to support its East Coast integrated steel plants [RESEARCHED]. The current procurement model relies heavily on daily, reactive spot chartering in highly volatile global dry bulk markets [PS]. This manual approach leads to missed cost-saving opportunities, improper vessel selection, and vessel idle time caused by strict physical draft, Length Overall (LOA), and beam limitations across East Coast Indian ports (Paradip, Vizag, Gangavaram, Gopalpur, Dhamra, Sagar-Sandheads, and Haldia) [PS].

To solve this, the proposed solution—NaviFreight-AI—is a predictive dry bulk chartering decision support system. It combines hybrid machine learning time-series forecasting (XGBoost/LightGBM + Temporal Fusion Transformer) [INFERENCE] with a Mixed-Integer Linear Programming (MILP) constraint solver [INFERENCE] and discrete-event port congestion simulation [INFERENCE]. NaviFreight-AI shifts SAIL from reactive single spot contracts to proactive short-to-medium term Contract of Affreightment (COA) / time-charter strategy [PS].

Fundamental System Objective:
To transition SAIL from reactive daily spot vessel chartering to an automated, predictive short-to-medium term chartering strategy that optimizes market entry timing, matches cargo volumes to port-compliant vessel classes (Capesize, Panamax, Supramax, Handysize), minimizes vessel idle/demurrage time, and mitigates freight market volatility risks [PS].

2. Problem Deconstruction

Below is the atomic deconstruction of explicit requirements extracted directly from the problem statement text, alongside inferred system implications and confidence tags.

3. Stakeholders & User Needs

4. Requirements Traceability

5. Constraints & Environment Realities

1. Geographical & Trade Route Realities:
   • Australia (Hay Point / Gladstone / DBCT) to East Coast India: ~20–25 sailing days [RESEARCHED].
   • USA (Hampton Roads / Baltimore) to East Coast India: ~40–45 sailing days via Cape of Good Hope / Suez [RESEARCHED].
   • Mozambique (Nacala / Beira) to East Coast India: ~14–18 sailing days [RESEARCHED].
   • Indonesia (Samarinda / Taboneo) to East Coast India: ~8–12 sailing days [RESEARCHED].
2. East Coast Port Physical Draft & Infrastructure Disparities:
   • Haldia Dock Complex (HDC): Riverine port on Hooghly river with severe draft constraints (~7.0m–8.5m tidal dependent) [RESEARCHED]. Cannot dock fully laden Supramax/Panamax. Requires transshipment/lighterage at Sagar-Sandheads anchorage [RESEARCHED].
   • Paradip Port: Deepwater port (~16.5m draft at coal berths) capable of handling Capesize and Panamax vessels [RESEARCHED].
   • Gangavaram Port: Deep draft (~19.5m draft) capable of handling Capesize and VLOC [RESEARCHED].
   • Vizag (Visakhapatnam): Outer harbor deep draft (~16.5m Capesize), Inner harbor constrained (~12.5m - 14.5m) [RESEARCHED].
   • Dhamra Port: Deepwater (~17.5m draft), Capesize compliant [RESEARCHED].
   • Gopalpur Port: Medium draft (~12.5m–14.5m), Panamax/Supramax compliant [RESEARCHED].
3. PSU Enterprise & Connectivity Environment:
   • Strict Cybersecurity: SAIL intranet environments enforce strict air-gapping, proxy firewalls, and data governance policies. External cloud tools must operate over secure HTTPS REST APIs with OAuth2/JWT authentication [INFERENCE].
   • Audit Compliance: Public Sector Undertakings (PSUs) in India must adhere to CVC (Central Vigilance Commission) guidelines. Every procurement recommendation must be mathematically explainable and logged for audit trails [KNOWN].

6. Research Questions

1. RQ-1: What open or commercial datasets best approximate Baltic Exchange dry bulk spot rates (BCI, BPI, BSI, BHSI) and specific coal trade routes (C5 Australia-China/Asia, P5 Indonesia) for hackathon prototyping?
2. RQ-2: What are the exact draft, LOA, beam, and unloader MT/day handling rate constraints for all 7 target East Coast Indian ports and foreign loading ports?
3. RQ-3: How can lighterage/transshipment at Sagar-Sandheads be mathematically modeled when evaluating deep-draft Capesize vessels destined for shallow Haldia/Durgapur supply chains?
4. RQ-4: What feature sets (bunker fuel VLSFO, metallurgical coal TSI, steel demand, AIS vessel density) yield the highest directional accuracy in 30-day and 60-day dry bulk freight rate forecasting?
5. RQ-5: How can a 4-6 student team simulate real-time vessel movement and port congestion without paying thousands of dollars for enterprise AIS streaming APIs?

7. Data Requirements & Sources

8. Existing Solutions & Gap Analysis

9. Candidate Solution Approaches

Approach 1: Deterministic Rule-Based Heuristic Solver + Statistical ARIMA/GARCH Forecasting

• Mechanism: Statistical ARIMA time-series models for spot rate forecasting combined with hardcoded rule-based lookup tables for vessel-port physical matching.
• Complexity: Low.
• Explainability: Very High (standard statistical formulas).
• Pros: Fast implementation, highly transparent for PSU audits.
• Cons: Fails to capture non-linear market shocks, geopolitical events, or complex multi-variable interactions (bunker prices + coal demand + port queueing).

Approach 2: Deep Learning (LSTM / PatchTST) + Mixed Integer Linear Programming (MILP)

• Mechanism: Deep Recurrent Neural Networks (LSTM/GRU) or Transformer-based time series models (PatchTST) to predict multi-route freight indices, coupled with Pyomo/OR-Tools MILP solver for vessel allocation.
• Complexity: High.
• Explainability: Medium-Low (Deep learning black-box requires SHAP/LIME explanation).
• Pros: Captures complex non-linear temporal trends.
• Cons: Requires massive historical training datasets; prone to overfitting during black-swan freight market shocks.

Approach 3 (RECOMMENDED): Hybrid Machine Learning (Ensemble LightGBM/XGBoost + Temporal Fusion Transformer) + MILP Constraint Solver + Discrete Event Congestion Simulator

• Mechanism: Ensemble ML (LightGBM + Prophet/TFT) incorporating exogenous indicators (bunker prices, coal demand, seasonal weather indices) to forecast freight rates with prediction confidence bounds. Physical port constraints, parcel splitting (lighterage at Sandheads), and contract selection (Spot vs. 6-month COA) are solved using Mixed-Integer Linear Programming (PuLP/OR-Tools). Real-time queueing and idle time risk are calculated via a discrete-event SIMPY queue simulator.
• Complexity: Medium-High.
• Explainability: High (Feature importance SHAP scores + explicit mathematical MILP constraints).
• Pros: High predictive accuracy, transparent decision logic, realistic port handling modeling, robust against noisy data.
• Cons: Requires integration of multiple ML and optimization libraries.

10. Recommended Approach & Why

The recommended approach is Approach 3: Hybrid ML Forecasting Engine + MILP Constraint Solver + Discrete Event Congestion Simulator.

Why Alternatives Were Rejected:

• Approach 1 Rejected: Pure statistical models (ARIMA) cannot adapt to violent freight market swings caused by global geopolitical disruptions, monsoon seasonality, or rapid fuel price spikes [INFERENCE].
• Approach 2 Rejected: Standalone deep learning models lack physical port constraint awareness and risk failing PSU audit standards due to "black-box" decisions without explicit constraint validation [INFERENCE].

Why Approach 3 Wins:

1. Separation of Forecasting & Decision Logic: Machine Learning is used where it excels (predicting volatile price trends from noisy multi-source data), while Mathematical Optimization (MILP) is used where strict compliance is required (draft, LOA, beam, parcel size hard constraints) [INFERENCE].
2. PSU Audit Compliance: The MILP mathematical formulation produces deterministic, auditable proofs for why a specific vessel class (e.g., Panamax vs. Capesize lightered at Sandheads) was selected over another [INFERENCE].
3. Hackathon Feasibility: Can be developed within the timeline using open-source Python frameworks (lightgbm, pulp, simpy, scikit-learn, FastAPI, React) [INFERENCE].

11. Technical Architecture

Layered Architecture Diagram

+-----------------------------------------------------------------------------------+
|                            USER INTERFACE LAYER (React + Vite)                    |
|  [ Cargo Request Portal ]  [ Market Entry Timing ]  [ Vessel Optimizer ]  [ Alerts ]|
+-----------------------------------------------------------------------------------+
                                         |  HTTPS / REST / WebSocket
+-----------------------------------------------------------------------------------+
|                            API GATEWAY & BACKEND (FastAPI)                        |
|  [ Auth & RBAC ]  [ Scenario Controller ]  [ Audit Logger ]  [ Offline Sync ]      |
+-----------------------------------------------------------------------------------+
         |                                |                               |
+------------------+             +------------------+             +------------------+
| FORECAST ENGINE  |             | OPTIMIZER ENGINE |             | SIMULATION ENGINE|
| (LightGBM/Prophet|             |  (PuLP / MILP)   |             | (SimPy / Queues) |
| Multi-Horizon ML)|             | Port/Vessel Rules|             | Demurrage Risk   |
+------------------+             +------------------+             +------------------+
         |                                |                               |
+-----------------------------------------------------------------------------------+
|                             DATA & STORAGE LAYER                                  |
|  [ PostgreSQL (Relational) ]   [ Redis (Cache) ]   [ Vector / DuckDB (Analytics) ] |
+-----------------------------------------------------------------------------------+
                                         |  ETL Pipeline
+-----------------------------------------------------------------------------------+
|                         EXTERNAL DATA & SIMULATED FEEDS                           |
|  [ Baltic Freight Proxies ]  [ Port Draft Matrix ]  [ VLSFO Bunker ]  [ AIS Stream]|
+-----------------------------------------------------------------------------------+

Detailed Data-Flow Narrative

1. Ingestion Step: Automated background ETL scripts fetch daily Baltic freight proxies, VLSFO bunker fuel prices, metallurgical coal indices, and AIS vessel queue updates into PostgreSQL/DuckDB [INFERENCE].
2. Forecast Step: The user inputs a cargo procurement request (e.g., 150,000 MT coking coal from Hay Point, Australia to East Coast India within a 45-day window). The LightGBM/Prophet ML engine generates daily predicted freight rates for Capesize (C5 route) and Panamax routes over the next 90 days, returning mean predictions along with P10/P90 confidence bands [INFERENCE].
3. Constraint & Optimization Step: The request and forecast curves are passed to the MILP Optimization Engine. The engine evaluates candidate vessel strategies:
   • Option A: Single Capesize vessel (150k MT) direct to deepwater port (Gangavaram or Paradip) [INFERENCE].
   • Option B: Single Capesize vessel with lighterage/top-off at Sagar-Sandheads anchorage down to 8.5m draft, then Haldia discharge [INFERENCE].
   • Option C: Two Panamax vessels (75k MT each) shipped directly to Vizag/Gopalpur [INFERENCE].
   • Option D: 6-Month Time Charter / COA contract vs. Spot market execution [INFERENCE].
4. Simulation Step: Selected routes run through the SimPy Discrete-Event Queue Simulator, using current port berth congestion levels to estimate laytime, waiting days, and expected demurrage costs ($/day) [INFERENCE].
5. Presentation Step: The FastAPI backend returns the optimal recommended charter window, vessel type, contract structure (COA vs Spot), total estimated cost breakdown ($/MT), and risk alerts to the React frontend [INFERENCE].

12. Technology Stack

13. AI/ML Justification

*Note: The Vessel Matching & Contract Optimization Engine explicitly uses Non-AI Mathematical Optimization (MILP), ensuring 100% adherence to physical draft, LOA, beam, and legal rules without black-box hallucination [INFERENCE].*

14. Data Architecture & Offline/Low-Network Design

Core Database Schema Entities (PostgreSQL)

[Port] 1---* [Berth]
  |
  +--- LOA_Max, Beam_Max, Draft_Max_Meters, Air_Draft_Max, Handling_Rate_TPD, Lighterage_Required_Flag

[Vessel_Class]
  |
  +--- Class_Name (Capesize/Panamax/Supramax/Handysize), DWT_Min, DWT_Max, Draft_Laden_Meters, LOA_Meters, Beam_Meters

[Freight_Route_Rate]
  |
  +--- Route_Code (C5/P5), Date, Spot_Rate_USD_MT, Time_Charter_Rate_USD_Day, Fuel_Bunker_Index

[Cargo_Procurement_Request]
  |
  +--- Request_ID, Cargo_Volume_MT, Origin_Port_ID, Destination_Port_ID, Target_Laycan_Start, Target_Laycan_End

[Optimization_Result]
  |
  +--- Scenario_ID, Recommended_Contract_Type, Recommended_Vessel_Class, Optimal_Market_Entry_Date, Total_Cost_USD_MT

Low-Network & Offline Resilience

• Progressive Web App (PWA): Frontend caches UI assets and static port constraint databases locally via Service Workers and IndexedDB [INFERENCE].
• Bandwidth Optimization: API responses compress JSON payloads using Gzip/Brotli, transmitting minimal differential payload updates for rate forecasts [INFERENCE].
• Background Synchronization: In field conditions with spotty network (e.g., dockside operations), local cargo requests are queued in IndexedDB and synchronized automatically upon reconnecting [INFERENCE].

15. Security, Privacy & Governance

• Data Privacy & Confidentiality: SAIL internal procurement volumes, target bid prices, and stockyard levels are classified as Proprietary Commercial Information [PS]. Inputs are encrypted using AES-256 at rest and TLS 1.3 in transit [INFERENCE].
• Role-Based Access Control (RBAC):
  • Admin: User management, system calibration.
  • Procurement Officer: Create requests, run forecasts, view contract recommendations.
  • Port Coordinator: Update live port draft/berth availability overrides.
  • Auditor / Executive: View historical decision logs, SHAP explainability reports, and cost savings metrics [INFERENCE].
• Governance & Audit Trail: To satisfy CVC/CAG public sector audit requirements, every optimization output records a immutable audit log capturing: timestamp, user ID, exact input parameters, active ML model version, MILP solver constraint status, and rationale output [INFERENCE].

16. MVP Scope (Hackathon Build)

In Scope (MUST BUILD)

• Interactive Cargo Procurement Portal: UI for logistics managers to input volume (e.g., 50k–180k MT), origin (Australia/US/Mozambique/Indonesia), destination (7 East Coast ports), and laycan window [PS].
• Port Infrastructure Constraint Matrix: Hardcoded, verifiable physical constraint rules for 7 East Coast Indian ports and 4 major overseas origin ports [PS].
• Multi-Route Freight Rate Forecasting Engine: LightGBM/Prophet model trained on proxy historical bulk index data predicting 15–60 day forward rates [PS].
• Vessel & Contract Strategy Optimizer (MILP): Recommends optimal vessel class (Capesize, Panamax, Supramax, Handysize), contract strategy (Spot vs. 6-Month COA), and market entry window [PS].
• Lighterage / Split Discharge Support: Specific economic modeling for Capesize lighterage at Sagar-Sandheads for Haldia-bound cargo [PS].
• Volatility Risk & Early Warning Dashboard: Visual alert system highlighting upcoming market volatility windows and port queue bottlenecks [PS].

Described Only / Out of Scope (DO NOT BUILD DURING HACKATHON)

• Real-time live satellite AIS streaming API subscriptions (simulated via local data generator) [ASSUMPTION].
• Direct SAP ERP two-way database integration with SAIL internal enterprise systems [INFERENCE].
• Automated contract execution or direct integration with shipping broker trading desks [INFERENCE].

17. Live Demo Simulation Plan

To ensure a flawless SIH demonstration without reliance on paid enterprise APIs, the team will implement a Deterministic Live Data Simulator:

1. Simulated AIS Vessel Tracker: Pre-scripted JSON generator simulating 15 bulk carriers navigating trade routes (e.g., MV Sail Coking traveling from Hay Point to Paradip) with realistic lat/long coordinates rendered on a Leaflet.js interactive map [INFERENCE].
2. Simulated Baltic Market Feed: Historical daily rate file appended with a mock "live market feed" socket push, allowing judges to test real-time forecast updates during live interaction [INFERENCE].
3. Scenario Testing Walkthrough for Judges:
   • Scenario 1 (Direct Capesize Deepwater): Input 150,000 MT coking coal from Hay Point to Paradip. System demonstrates Capesize recommendation, showing $4.50/MT freight savings over using two Panamax vessels [INFERENCE].
   • Scenario 2 (Shallow Port Lighterage): Input 120,000 MT bound for Haldia. System detects Haldia 8.0m draft restriction, automatically recommends Capesize lighterage at Sagar-Sandheads down to permissible draft before river transit [INFERENCE].
   • Scenario 3 (Market Rally Risk): System detects a simulated 20% spike in bunker fuel prices, generating an early warning alert recommending securing a 6-month COA immediately before spot rates surge [INFERENCE].

18. Implementation Roadmap

Alignment Deadline: 30 September 2026 [PS]

+-----------------------------------------------------------------------------------+
| PHASE 1: Data Pipeline & Baseline ML Models (Weeks 1 - 6)                         |
| - Curate Baltic dry bulk historical dataset & port constraint matrix              |
| - Train baseline Prophet & LightGBM time-series forecasting models               |
+-----------------------------------------------------------------------------------+
                                         |
+-----------------------------------------------------------------------------------+
| PHASE 2: MILP Optimizer & SimPy Congestion Simulator (Weeks 7 - 12)               |
| - Develop PuLP constraint solver for vessel-port physical matching                |
| - Implement Sandheads lighterage economic model & SimPy port queue simulator     |
+-----------------------------------------------------------------------------------+
                                         |
+-----------------------------------------------------------------------------------+
| PHASE 3: React Dashboard & Scenario Integration (Weeks 13 - 18)                   |
| - Build interactive frontend dashboard with Recharts & Leaflet GIS tracking map   |
| - Integrate FastAPI backend endpoints with RBAC & Audit Trail Logging             |
+-----------------------------------------------------------------------------------+
                                         |
+-----------------------------------------------------------------------------------+
| PHASE 4: Validation, Risk Testing & Demo Polish (Weeks 19 - 24 / Sept 2026)      |
| - Execute false-positive/negative forecast validation & security audit           |
| - Package live demo simulator and finalize SIH submission dossier                |
+-----------------------------------------------------------------------------------+

19. Testing & Validation Plan

20. Risk Register

21. Cost & Scalability

Hackathon MVP Build Cost: $0 (Free Tier Stack)

• Frontend Hosting: Vercel / Netlify Free Tier ($0).
• Backend Hosting: Render / Cloud Run Free Tier ($0).
• Database: PostgreSQL on Supabase Free Tier ($0).
• Data & APIs: Scraped/Open Proxy Datasets + Local Python Simulators ($0).

Production Enterprise Deployment (SAIL Enterprise Cloud)

• Cloud Infrastructure (AWS / Azure India Cloud): $12,000 – $18,000 / year (Managed PostgreSQL, Redis, Kubernetes nodes) [INFERENCE].
• Commercial Shipping Data Feeds (Baltic Exchange / AIS API): $25,000 – $40,000 / year [RESEARCHED].
• Scalability Potential: Architecture easily scales to handle all raw material maritime imports across SAIL plants (including iron ore, limestone, coking coal) and can be white-labeled for other Indian steel PSUs (e.g., RINL, NMDC) [INFERENCE].

22. Differentiation

23. Future Scope

1. Integration with Inland Rail Freight (FOIS): Link discharge port scheduling directly with Indian Railways' Freight Operations Information System (FOIS) for end-to-end rake movement to steel plants [INFERENCE].
2. Carbon Emission & EU ETS Tracking: Calculate CII (Carbon Intensity Indicator) scores and carbon offset costs for candidate charter vessels [INFERENCE].
3. Multi-Commodity Expansion: Extend optimization models to include imported limestone, dolomite, and iron ore coastal shipping [INFERENCE].

24. Assumptions & Open Items Log

Consolidated Assumptions [ASSUMPTION]

• [ASSUMPTION] Public bulk freight proxies (e.g., FRED dry bulk indices) provide sufficient feature correlation to prototype Baltic Exchange route dynamics for the hackathon MVP.
• [ASSUMPTION] Port draft and handling capacity figures published in official Port Trust master plans represent baseline operational constraints.
• [ASSUMPTION] Real-time AIS vessel streaming can be realistically simulated during the hackathon demonstration without impacting architectural validity.

Open Items & Verification Needs [NEEDS VALIDATION]

• [NEEDS VALIDATION] Exact current commercial tariff structure and lighterage rates ($/MT) at Sagar-Sandheads anchorage for Haldia-bound vessels.
• [NEEDS VALIDATION] Precise historical ratio of long-term COA contracts versus spot charters executed by SAIL chartering desk over the last 3 fiscal years.

Final Decision & Metrics

• Recommended Solution: NaviFreight-AI (Hybrid ML Forecasting + MILP Vessel-Port Constraint Solver + SimPy Congestion Simulator) [INFERENCE].
• Recommended Stack: React 18 + FastAPI + PostgreSQL + LightGBM/Prophet + PuLP MILP Solver + Leaflet.js [INFERENCE].
• Why This Approach Wins: Mathematically balances volatile freight price forecasting with zero-hallucination physical port constraint compliance and auditable COA contract strategy recommendations [INFERENCE].
• Biggest Dependency: Access to representative historical dry bulk trade route freight data [INFERENCE].
• Biggest Risk: Sudden geopolitical black-swan events distorting short-term ML time-series predictions [INFERENCE].
• Biggest Technical Challenge: Modeling multi-stage split discharge and lighterage economics at Sagar-Sandheads for shallow-draft riverine entry to Haldia [INFERENCE].
• First 5 Actions:
  1. Construct JSON physical constraint matrix for all 7 East Coast Indian ports and foreign origin ports [INFERENCE].
  2. Scrape and normalize 5-year historical dry bulk proxy rate data and VLSFO bunker prices [INFERENCE].
  3. Implement PuLP MILP solver baseline for vessel-port physical matching [INFERENCE].
  4. Train baseline LightGBM/Prophet rate forecasting model [INFERENCE].
  5. Build React UI mockups for Cargo Request Portal and Decision Dashboard [INFERENCE].
• Validation Before Coding: Verify draft limit edge cases with published Port Authority notifications [INFERENCE].
• Buildability Score: 92 / 100 (Highly feasible for a 4–6 student team using standard Python ML/Optimization tools and React) [INFERENCE].
• Confidence: HIGH [INFERENCE].

Research Sources

1. Title: India's SAIL to import trial coking coal cargo from Mongolia / Coking Coal Import Dependency [1.1, 1.2, 1.5]

   • URL: https://www.steelorbis.com/steel-news/latest-news/indias-sail-to-import-trial-coking-coal-cargo-from-mongolia-1391697.htm
   • Source Type: Industry News / Press Report
   • Supported Claim: SAIL coking coal import volume dependency, origins (Australia, US, Mozambique, Russia), and voyage durations (20–25 days Australia, 40–45 days US).
   • Reliability: High.

2. Title: Maritime India Vision 2030 / Paradip & Haldia Port Infrastructure Reports [1.4, 2.1, 2.3]

   • URL: https://wwwcdn.imo.org/localresources/en/OurWork/Environment/Documents/Air%20pollution/Maritime%20India%20vision%202030.pdf
   • Source Type: Government Report / Ministry of Ports, Shipping and Waterways
   • Supported Claim: Physical port specifications, Haldia riverine draft constraints, dredging requirement, lighterage at Sandheads, and Paradip/Gangavaram deepwater berth capabilities.
   • Reliability: Official Government Benchmark.

3. Title: Baltic Exchange Dry Services & Route Assessments [3.2, 3.5]

   • URL: https://www.balticexchange.com/en/data-services/market-information0/dry-services.html
   • Source Type: Official Shipping Standards Organization
   • Supported Claim: Definitions of BDI, BCI, BPI, BSI, BHSI indices, vessel DWT/LOA/draft parameters (Capesize 180k DWT, Panamax 82k DWT), and C5/P5 route weighting.
   • Reliability: Official Maritime Benchmark.

4. Title: Machine Learning in Freight Rate Forecasting: A Systematic Literature Review [4.4, 4.8]

   • URL: https://research.cbs.dk/en/publications/machine-learning-in-freight-rate-forecasting-a-systematic-literat/
   • Source Type: Academic Peer-Reviewed Publication
   • Supported Claim: Superior performance of hybrid ML models (Neural Networks / Gradient Boosting + Time Series) over standalone linear models in volatile dry bulk rate forecasting.
   • Reliability: Academic High.