ForecastBlend — Production-Ready Multi-Model Weather Intelligence Platform

ForecastBlend is a meteorological intelligence and multi-model ensemble platform that ingests, normalizes, evaluates, and dynamically blends predictions from global numerical weather prediction (NWP) models and cutting-edge machine learning weather models. It answers three essential questions for operational meteorology and data-driven planning:

1. What is the expected weather at my location?
2. How do competing forecast models disagree, and what does their synthesized blend indicate?
3. How reliable is the forecast, and what empirical verification evidence supports that assessment?

---

🌐 Live Production Deployment

• Interactive Dashboard: https://forecast-blend.vercel.app
• Production Serverless API: https://forecast-blend.vercel.app/api/v1/forecast/PUN
• API Health Check: https://forecast-blend.vercel.app/api/v1/health

---

📸 Dashboard Preview

Blended Forecast & Station Network	Multi-Model Horizon Comparison Matrix

Empirical Verification Lab & Skill Benchmarking

---

🏛️ System Architecture

ForecastBlend is designed with a strict, decoupled layered architecture separating raw weather ingestion, empirical skill verification, mathematical blending, and presentation.

graph TD
    subgraph Data Sources
        GFS[NOAA GFS 0.25° NWP]
        ECMWF[ECMWF IFS 0.1° NWP]
        GC[DeepMind GraphCast AI 0.25°]
        OBS[IMD & Synoptic In-situ Ground Observations]
    end

    subgraph Backend / Engine [Node.js & TypeScript Server]
        STORE[In-Memory / FS Data Store]
        SCORER[Skill Scorer: Rolling 14-Day MAE / RMSE]
        WEIGHTS[Regularized Inverse-MAE Softmax + Laplace Shrinkage]
        BLENDER[Ensemble Blender: Weighted Mean + Spread + Agreement]
        EXTREME[IMD Calibrated Extreme Weather Detector]
        INSIGHTS[Rule-Based Data-Grounded Meteorological Insights]
        API[Express REST API /api/v1]
    end

    subgraph Frontend [React 18 + TypeScript + Vite]
        NAV[Brand & UTC Pipeline Status]
        LOC[Location Bar & Station Geographic Metadata]
        HERO[Blended Forecast Hero + 95% Uncertainty Bounds]
        MATRIX[Comparative Multi-Model Horizon Matrix + Chart]
        VERIF[Verification Lab: Empirical Skill & Error Reduction]
        PANEL[Weather Insights Panel with Evidence Citations]
        ALERTS[IMD Risk Indicators & Safety Advisory Stream]
        CONSOLE[7-Stage Pipeline Execution Console]
    end

    GFS --> STORE
    ECMWF --> STORE
    GC --> STORE
    OBS --> STORE

    STORE --> SCORER
    SCORER --> WEIGHTS
    WEIGHTS --> BLENDER
    BLENDER --> EXTREME
    BLENDER --> INSIGHTS
    STORE --> API
    BLENDER --> API
    EXTREME --> API
    INSIGHTS --> API

    API --> NAV
    API --> LOC
    API --> HERO
    API --> MATRIX
    API --> VERIF
    API --> PANEL
    API --> ALERTS
    API --> CONSOLE

Loading

---

📐 Scientific Blending Methodology

Rather than treating model consensus as an unweighted average or fabricating confidence scores, ForecastBlend implements a regularized inverse-error weighting scheme grounded in statistical ensemble forecasting.

1. Rolling Skill Evaluation (MAE)

For each model $m \in \mathcal{M}$, geographic region $r$, lead time $h \in {24, 48, 72, 96}$, variable $v$, and weather regime $\omega$:

$$\text{MAE}_{m, r, h, v, \omega} = \frac{1}{N} \sum_{k=1}^N \left| \hat{y}_{m, k} - y_{\text{obs}, k} \right|$$

where $N$ is the sample count over a rolling 14-day verification window. Lookahead bias is strictly prevented: observations evaluated were recorded strictly before the forecast issue time.

2. Regularized Inverse-MAE Softmax Weighting with Laplace Shrinkage

To prevent single-model overfitting and guard against division by zero, raw skill weights are computed using an epsilon regularizer $\epsilon = 0.05$:

$$S_m = \frac{1}{\text{MAE}_m + \epsilon}$$

Normalized initial probabilities:

$$p_m = \frac{S_m}{\sum_{j=1}^M S_j}$$

To prevent extreme weight swings when sample sizes are small or when an individual model experienced temporary anomalies, Laplace Shrinkage is applied toward a uniform uninformative prior ($1/M$):

$$\hat{w}_m = (1 - \lambda) \cdot p_m + \lambda \cdot \frac{1}{M}$$

where $\lambda = 0.15$. The resulting weights are strictly positive and satisfy:

$$\sum_{m \in \mathcal{M}} \hat{w}_m = 1.0$$

3. Blended Expected Value & Physical Bounds

$$\bar{y}_{\text{blend}} = \sum_{m \in \mathcal{M}} \hat{w}_m \cdot \hat{y}_m$$

Physical non-negativity constraints are strictly enforced:

$$\bar{y}_{\text{blend}} = \max\left(0, \bar{y}_{\text{blend}}\right) \quad \text{for } v \in {\text{rainfall_mm}, \text{wind_kmh}}$$

4. Ensemble Spread ($\sigma$) and Model Agreement Index ($A$)

Ensemble variance is calculated as the weighted sample dispersion:

$$\sigma = \sqrt{\sum_{m \in \mathcal{M}} \hat{w}_m \left(\hat{y}_m - \bar{y}_{\text{blend}}\right)^2}$$

The Model Agreement Index $A \in [0, 1]$ measures inter-model consensus relative to domain-specific dispersion scales $\sigma_{\text{ref}}$ (Rainfall: 8.0 mm, Temperature: 2.5 °C, Wind: 6.0 km/h):

$$A = \exp\left(-\frac{\sigma}{\sigma_{\text{ref}}}\right)$$

• $A \to 1.0$: Tight clustering between global models (high consensus).
• $A \to 0.0$: High inter-model divergence (synoptic uncertainty).

5. Statistical 95% Uncertainty Interval

Assuming an asymptotic normal error distribution around the ensemble mean:

$$I_{95%} = \left[ \max\left(0, \bar{y}_{\text{blend}} - 1.96\sigma\right), ; \bar{y}_{\text{blend}} + 1.96\sigma \right]$$

---

⚠️ IMD-Calibrated Meteorological Indicators

ForecastBlend detects significant regional weather risks using official thresholds defined by the India Meteorological Department (IMD):

Category	Variable	Threshold	IMD Classification	Severity
Precipitation	rainfall_mm	$\ge 64.5\text{ mm / 24h}$	Heavy Rainfall	Watch
Precipitation	rainfall_mm	$\ge 115.6\text{ mm / 24h}$	Very Heavy Rainfall	Warning
Precipitation	rainfall_mm	$\ge 204.5\text{ mm / 24h}$	Extremely Heavy Rainfall	Severe
Heatwave	temp_c	$\ge 40.0^\circ\text{C}$	Heatwave Advisory	Watch
Heatwave	temp_c	$\ge 43.0^\circ\text{C}$	Heatwave Condition	Warning
Heatwave	temp_c	$\ge 45.0^\circ\text{C}$	Severe Heatwave Condition	Severe
Gale Wind	wind_kmh	$\ge 50.0\text{ km/h}$	Gusty Wind Watch	Watch
Gale Wind	wind_kmh	$\ge 62.0\text{ km/h}$	Gale Wind Warning	Warning
Gale Wind	wind_kmh	$\ge 88.0\text{ km/h}$	Severe Gale Condition	Severe

Safety Advisory Notice: Indicators generated by ForecastBlend are rule-based decision support tools derived from numerical model blending and do not constitute official civil protection warnings. Always consult the India Meteorological Department for official life and property safety directives.

---

🗺️ Meteorological Stations & Coverage

ForecastBlend monitors high-impact synoptic regions in Western and Central India:

Station Code	City	Latitude	Longitude	Elevation	Climate Regime
PUN	Pune	18.5204° N	73.8567° E	560 m	Semi-arid / Western Ghats Rain Shadow
MUM	Mumbai	19.0760° N	72.8777° E	14 m	Tropical Wet & Dry / Coastal Maritime
NGP	Nagpur	21.1458° N	79.0882° E	310 m	Tropical Wet & Dry / Central Continental

---

🤖 Contributing Weather Models

ForecastBlend combines three structurally distinct forecasting paradigms:

1. NOAA Global Forecast System (GFS)
   • Type: Hydrostatic Primitive Equation Global NWP
   • Resolution: 0.25° (~22 km), 6-hour cycles (00Z, 06Z, 12Z, 18Z)
   • Characteristics: High skill at extended horizons (+72h to +96h); known orographic precipitation dispersion bias in steep coastal topography.
2. ECMWF Integrated Forecasting System (IFS)
   • Type: Non-hydrostatic Spectral Global NWP
   • Resolution: 0.1° (~9 km), 12-hour cycles (00Z, 12Z)
   • Characteristics: Gold-standard synoptic precipitation accuracy; conservative peak temperature estimations in convective regimes.
3. DeepMind GraphCast AI Weather Model
   • Type: Autoregressive Graph Neural Network (GNN) on Icosahedral Mesh
   • Resolution: 0.25° (~25 km), trained on 40 years of ECMWF ERA5 reanalysis
   • Characteristics: Extremely fast inference (<60s); superior mid-troposphere temperature tracking; smoother spatial precipitation gradients.

---

⚡ REST API Reference

All endpoints return JSON and include security headers (X-Content-Type-Options: nosniff, CORS enabled).

1. Multi-Model Blended Forecast

GET /api/v1/forecast/:region

{
  "region": "PUN",
  "issued_at": "2026-10-22T06:00:00.000Z",
  "location": {
    "code": "PUN",
    "name": "Pune",
    "latitude": 18.5204,
    "longitude": 73.8567,
    "elevationM": 560,
    "climateZone": "Semi-arid / Western Ghats Rain Shadow"
  },
  "current_observation": {
    "region": "PUN",
    "observedAt": "2026-10-22T06:00:00.000Z",
    "temp_c": 28.5,
    "rainfall_mm": 0.0,
    "wind_kmh": 14.0,
    "source": "In-situ Synoptic Weather Station & Reanalysis Baseline"
  },
  "forecast": {
    "rainfall_mm": {
      "24": {
        "value": 14.2,
        "confidence": 0.782,
        "spread": 2.14,
        "uncertaintyRange": { "min": 10.01, "max": 18.39 },
        "weights": { "NWP_GFS": 0.28, "NWP_ECMWF": 0.44, "AI_GRAPHCAST": 0.28 },
        "models": { "NWP_GFS": 12.0, "NWP_ECMWF": 15.5, "AI_GRAPHCAST": 14.8 }
      }
    }
  }
}

2. Other Key Endpoints

Method	Endpoint	Description
GET	/api/v1/locations	Returns all active meteorological stations with geographic metadata
GET	/api/v1/alerts?region=MUM	Returns active IMD-calibrated extreme weather indicators
GET	/api/v1/insights/:region	Rule-based meteorological insights with supporting model evidence
GET	/api/v1/verification/:region	Empirical 14-day rolling MAE/RMSE comparisons and error reduction %
GET	/api/v1/weights/:region	Active regularized model weights across horizons and weather regimes
GET	/api/v1/health	Service uptime, version, and operational health status

---

🛠️ Local Development & Quickstart

Prerequisites

• Node.js 18.x or 20.x
• npm 9.x or later

Installation

# Clone the repository
git clone https://github.com/Avadhut-Pawar31/ForecastBlend.git
cd ForecastBlend

# Install root, server, and dashboard dependencies
npm install

# Copy environment example
cp .env.example .env

Running Locally

# Start both Dashboard (Vite) and Server (Express) concurrently:
npm run dev:server     # Runs server on http://localhost:3001
npm run dev:dashboard  # Runs dashboard on http://localhost:5173

Running Tests & Production Build

# Execute Vitest test suite (Algorithms, Extreme Detector, API Integration)
npm test

# Build production bundles (TypeScript compilation & Vite bundle)
npm run build

---

🧪 Automated Testing Suite

ForecastBlend includes comprehensive unit and integration tests using Vitest:

• Weight Normalization & Shrinkage: Validates that regularized weights strictly sum to 1.0, favor lower MAE models, and apply Laplace shrinkage toward equal priors on missing model data.
• Blending & Bounds: Validates weighted expectations, non-negative precipitation clamping, ensemble standard deviations, and Model Agreement Indices.
• Extreme Event Detection: Tests IMD thresholds for heavy rainfall, heatwaves, and gale winds with citations.
• Data-Grounded Insights: Verifies evidence generation from multi-model forecasts.
• HTTP API Integration: Tests status codes (200, 400, 404), JSON schema compliance, and parameter validation.

Run tests anytime with:

npm --prefix server run test

---

📄 License

This project is licensed under the MIT License — see the LICENSE file for details.

Developed with passion for open meteorological science and reliable weather intelligence.