Renewable Energy

A large renewable energy company like Adani Green manages 20+ GW across hundreds of plants. Architecture: 1) Data Ingestion: Each plant has SCADA system (local) collecting: inverter data (AC power, DC voltage/current, temperature, fault codes), weather station (irradiance GHI/POA, temperature, wind speed), meter data (grid export, import), tracker position (for single-axis trackers). Data volume: 500 MW plant → 200+ inverters → 2,000+ data points → every 5 seconds = 17M data points/day per plant. 50 plants = 850M data points/day. Protocol: Modbus from field devices → plant SCADA (OPC-UA) → cloud via MQTT/AMQP. 2) Cloud Architecture: Data landing: IoT Hub (Azure) or IoT Core (AWS) — handles device connectivity, authentication, message routing. Stream processing: Kafka → Spark Streaming for real-time metrics (plant generation, PR, availability). Batch processing: hourly/daily aggregation, performance ratio calculation, loss analysis. Storage: TimescaleDB for time-series (hot: 90 days), S3/Blob for cold storage (raw data: 25 years for PPA duration). 3) Analytics: a) Performance Ratio (PR): PR = Actual Energy / (POA Irradiance × STC Capacity × Time). Target: 80-85%. Break down losses: temperature (5-8%), soiling (2-5%), inverter (2-3%), clipping (1-2%), shading, curtailment. b) Fleet benchmarking: compare PR across plants — control for irradiance and temperature. Identify underperformers. c) Predictive maintenance: inverter failure prediction — train ML model on: temperature cycling, DC/AC ratio drift, MPPT efficiency decline. Predict failure 2-4 weeks ahead. Drone thermal inspection: computer vision detects hot spots on modules (failed bypass diode, cracked cell). d) Forecasting: per-plant and portfolio-level generation forecast. Ensemble ML model. Used for: grid scheduling (avoid deviation penalties), revenue forecasting (financial planning). 4) Business Intelligence: Executive dashboard: portfolio generation (GWh/day), revenue (₹ crore/month), PR trend, availability, curtailment. Plant manager: detailed plant view, alarm management, maintenance scheduling. Financial: actual vs P50/P90 generation comparison, PPA billing, REC generation.

Grid-scale BESS (100 MW+ installations) must be optimized for revenue, grid services, and battery longevity. Architecture: 1) Battery Management: Hardware BMS (Battery Management System) monitors every cell: voltage (3.2-3.6V for LFP), temperature, current. Ensures: no cell overcharged, no cell over-discharged, thermal limits respected. Software BMS: aggregates cell data → module → rack → container → plant level. State of Charge (SoC): calculated from coulomb counting + OCV (Open Circuit Voltage) calibration. State of Health (SoH): measured capacity vs nameplate — tracks degradation over time. 2) Degradation Model: Battery degrades with: a) Cycle aging: each charge/discharge cycle reduces capacity. Cycles at high DoD (Depth of Discharge) degrade more. b) Calendar aging: degradation over time even without cycling. c) Temperature: high temperature accelerates degradation exponentially (Arrhenius). d) C-rate: fast charging/discharging degrades more. Model: empirical model trained on manufacturer data + field data. Predicts: remaining useful life (years), remaining capacity (MWh). Warranty typically: 80% capacity at 10-15 years. Optimization must balance: revenue (cycle more = more revenue) vs degradation (cycle more = shorter life). 3) Revenue Optimization: Multiple revenue streams: a) Energy arbitrage: charge when electricity cheap (solar hours: ₹2-3/kWh), discharge when expensive (evening peak: ₹8-12/kWh). Margin = ₹5-10/kWh × daily cycles. b) Frequency regulation: respond to grid frequency deviations within 100ms. CERC ancillary services market. Revenue: ₹15-20/MW/hour for regulation capability. c) Solar/wind firming: pair with renewable — store surplus, deliver when needed. Converts variable renewable into firm (dispatchable) power. Premium: ₹1-2/kWh over variable renewable tariff. d) Peak shaving: industrial consumers reduce peak demand charges. Optimization algorithm: Mixed Integer Linear Programming (MILP). Inputs: price forecast, generation forecast (if paired with solar), degradation model, grid service commitments. Output: optimal charge/discharge schedule for each 15-minute block. Re-optimized every hour as conditions change. 4) Safety: Thermal runaway risk: lithium-ion cells can catch fire if overcharged/overheated. Detection: gas sensors (detect off-gassing before thermal runaway), cell temperature monitoring, voltage anomaly detection. Suppression: aerosol or gas-based fire suppression in each container. Separation: fire walls between containers. LFP (Lithium Iron Phosphate) preferred over NMC for grid-scale due to better thermal stability.

Renewable forecasting is critical because: 1) Grid stability: grid must balance supply=demand every second. Variable renewables make this harder. Better forecasts → grid operator can plan reserves efficiently. 2) Market scheduling: CERC mandates day-ahead scheduling for renewables. Deviation penalties: ₹0.5-2/kWh for forecast errors beyond ±15%. For a 500 MW solar plant generating 2.5 GWh/day, 1% forecast improvement = ₹5-10 lakh/day savings. 3) Storage optimization: when to charge/discharge BESS depends on renewable forecast. ML Model Architecture: a) Features: Weather (primary driver): NWP model output — irradiance (GHI, DNI, DHI), temperature, cloud cover, wind speed/direction. Multiple NWP models: GFS (free, 12-hour update), ECMWF (paid, best accuracy), WRF (regional, customizable). Satellite: INSAT imagery processed every 15 min — current cloud cover, cloud motion vectors (1-4 hour nowcast). Historical: past generation data, past weather, seasonal patterns. Calendar: month, day of year (sun angle), time of day. Plant-specific: installed capacity, tilt/azimuth, tracker type, soiling level, scheduled maintenance. b) Model Architecture: Ensemble approach (not single model): i) Physical model: irradiance → module temperature → DC power → AC power (pvlib/PVsyst equations). Good for clear days, handles equipment changes. ii) Statistical: ARIMA/SARIMAX — captures autocorrelation in generation time series. Good for very short-term (1-4 hours). iii) ML: Gradient Boosting (XGBoost/LightGBM) on weather + calendar features. Best for day-ahead. iv) Deep Learning: LSTM/Transformer network — sequential weather data → generation sequence. Best for capturing weather regime changes. Ensemble: weighted average of all models — weights optimized via cross-validation. c) Training: 2+ years of concurrent generation + weather data. Cross-validation: walk-forward (train on past, test on next month — never use future data). Metrics: MAPE (Mean Absolute Percentage Error), RMSE, forecast skill score (vs persistence model). d) Deployment: Automated pipeline: NWP data downloaded (6 AM, 6 PM) → feature engineering → model inference → schedule generation → submit to SLDC. Monitoring: track forecast accuracy daily, retrain monthly with latest data. Human override: forecaster can adjust during extreme weather (cyclone, heat wave).

Carbon credits represent avoided or removed greenhouse gas emissions. Technology Architecture: 1) How Carbon Credits Work: One credit = 1 tonne CO₂ equivalent (tCO₂e) avoided or removed. Example: 100 MW solar plant generates 200 GWh/year → avoids 160,000 tCO₂e (India grid emission factor: 0.8 tCO₂/MWh). These 160,000 credits can be sold to companies wanting to offset their emissions. Price: $5-15/tonne (voluntary market), potentially higher in compliance markets. Revenue: $0.8-2.4M/year additional income for the solar plant. 2) MRV (Measurement, Reporting, Verification): Technology backbone of carbon markets — ensures credits represent real emission reductions. Measurement: continuous monitoring of generation (solar SCADA), baseline calculation (what emissions would have occurred without the project — typically grid electricity). Reporting: annual monitoring report with generation data, emission calculations, methodology adherence. Verification: independent third-party auditor (DOE — Designated Operational Entity) audits the report. Technology: automated data collection from plant monitoring → carbon calculation engine (applies UNFCCC-approved methodology) → generate monitoring report → submit to registry. 3) Registry Technology: Verra (VCS), Gold Standard, Clean Development Mechanism (CDM) — each maintains a registry database. Credit lifecycle: issuance (after verification) → trading → retirement (used for offset, can't be resold). Blockchain emerging: some registries use blockchain for immutable tracking — prevent double counting (same credit sold twice). India's CCTS (Carbon Credit Trading Scheme): new compliance market under Energy Conservation Act. BEE (Bureau of Energy Efficiency) administers. Will create domestic carbon market — obligated entities must hold credits. 4) REC Market (India-specific): Renewable Energy Certificate — 1 MWh renewable generation = 1 REC. Traded on IEX (Indian Energy Exchange). Obligated entities (DISCOMs, large consumers) must meet RPO (Renewable Purchase Obligation) — buy RECs if own renewable generation is insufficient. Price: ₹1,000-3,000 per REC (market-determined). Technology: generation data from RE plant SCADA → verified by state nodal agency → REC issued on IEX platform → traded in monthly sessions → retired against RPO obligation. 5) ESG Reporting: Companies report emissions under: BRSR (Business Responsibility and Sustainability Reporting — SEBI mandate), CDP (Carbon Disclosure Project), TCFD (Task Force on Climate-related Financial Disclosures). Technology platform: collect Scope 1 (direct emissions), Scope 2 (electricity), Scope 3 (supply chain) data → calculate carbon footprint → track against targets → report to stakeholders.

Solar + wind + battery storage combination is becoming the dominant new power model. It converts variable renewable into 'firm' (reliable, schedulable) power that can compete with coal plants. Design: 1) Plant Configuration (Hybrid): Example: 300 MW solar + 100 MW wind + 100 MW/400 MWh battery storage. Why hybrid? Solar peaks midday, wind often stronger at night/monsoon. Complementary profiles reduce total variability. Battery handles remaining gaps. PPA commitment: deliver 200 MW firm power for 12 hours/day (6 AM - 6 PM). 2) Sizing: Solar sizing: oversized relative to PPA — 300 MW solar for 200 MW PPA. During peak solar, excess charges battery. Rule of thumb: 1.5-2x RE capacity vs firm commitment. Battery sizing: Must bridge gaps. Worst case: 3 hours of low solar (cloud cover). 200 MW × 3 hours = 600 MWh needed. With battery degradation margin: 100 MW / 400 MWh (4-hour duration). Energy balance simulation: model 8,760 hours with historical weather → verify PPA commitment met 95%+ of the time. 3) Control System Architecture: Central controller (Energy Management System): every 100ms, decides: a) How much solar/wind to export directly to grid? b) How much to store in battery? c) How much to discharge from battery? d) Any curtailment needed (grid constraint)? Logic: If RE generation > PPA commitment → charge battery (limited by SoC and C-rate). If RE generation < PPA commitment → discharge battery to fill gap. If battery SoC low and RE generation low → potential PPA shortfall (penalty). Predictive: use forecast to pre-position battery SoC. If tomorrow expects clouds in afternoon, charge battery to 100% by noon. 4) Optimization Layers: Layer 1 (Real-time): dispatch controller — every second, maintain grid commitment, respond to frequency. Layer 2 (15-minute): schedule optimizer — optimal charge/discharge for next block based on updated forecast. Layer 3 (Day-ahead): strategic optimizer — battery SoC trajectory for next 24 hours, considering: weather forecast, market prices, degradation cost. 5) Economics: Solar: ₹2.5/kWh (levelized). Wind: ₹3.0/kWh. Battery: ₹7-10/kWh (levelized for stored energy — but improving fast as costs drop). Blended firm power: ₹4-5/kWh — competitive with new coal (₹4.5-5.5/kWh). As battery costs fall 10-15%/year, firm RE will be cheaper than coal by 2026-27. This is why utilities worldwide are canceling new coal plants.