Rooftop Solar PV Microgrid Sizing & ML Yield Forecasting

This repository contains a feasibility study and machine learning forecasting pipeline initiated in late 2023 to model the future rooftop solar photovoltaic (PV) microgrid at the École Centrale Casablanca campus in Bouskoura, Morocco (coordinates: 33.4806° N, -7.6183° W).

The study models the grid's performance for the target operational year of 2025 by establishing a realistic physical system layout through standard flat-roof geometric spacing and thermodynamic thermal-loss constraints. It trains an XGBoost model on real-world PV SCADA telemetry to emulate hardware performance, and transposes those learned physical relationships to Casablanca’s actual weather to generate accurate, loss-adjusted generation forecasts.

---

Project Deliverables

• Full Technical Report: Read our comprehensive engineering feasibility study in report.pdf.
• Interactive Notebook: Explore the complete data cleaning, anomaly detection, and machine learning pipeline in notebooks/campus-energy-optimization.ipynb.

---

1. Project Context & Objectives

In 2023, it was estimated that the university campus consumes $1,546,749\text{ kWh/year}$. Following a facilities and residential expansion plan, the annual demand was projected to rise to $2,162,213\text{ kWh/year}$. This study evaluates the maximum feasible on-site rooftop PV generation potential across 35 residential pavilions, the main Pedagogical Buildings, and the Galilée Building to offset the campus's rising utility costs and carbon footprint.

---

2. Physical System Sizing & Layout

In order to resolve the physical limitations of rooftop arrays, we implemented standard flat-roof layout constraints:

• 50% Spacing Rule: Flat roofs require tilted panels (fixed at $30^\circ$ South-facing to optimize winter generation). To prevent tilted rows from casting shadows on one another, a flat-roof utilization factor of $50%$ was applied.
• System Sizing Breakdown:
  • 35 Pavilions: $35 \times 70\text{ m}^2 \text{ roof area} = 2,450\text{ m}^2$
  • Pedagogical Buildings: $7,000\text{ m}^2 \text{ roof area}$
  • Galilée Building: $800\text{ m}^2 \text{ roof area}$
  • Total Usable Roof Area: $10,250\text{ m}^2 \times 0.50 = \mathbf{5,125\text{ m}^2}$
• Module Selection: Tesla Solar Plasma TS545PR ($545\text{ Wp}$, Mono PERC, $2.58\text{ m}^2$ per panel)
• Total Physical Capacity: $$\text{Total Modules} = \frac{5,125\text{ m}^2}{2.58\text{ m}^2} \approx 1,986\text{ panels}$$ $$\text{Total Capacity} = 1,986 \text{ panels} \times 0.545\text{ kWp} = \mathbf{1,082.37\text{ kWp}}$$

---

3. ML Pipeline

Our forecasting model was trained on a real-world PV SCADA database (Rostamipour & Ghayeni 2026) consisting of 90 days of continuous 30-minute inverter telemetry and collocated weather station readings.

Data Processing Steps:

1. Temperature Unit Mismatch: logging mismatches were found where one weather station switched outputs from Celsius to Fahrenheit ($^\circ\text{F}$), creating outliers. These were isolated and corrected to Celsius.
2. Electrical Scaling Glitches: Some inverters logged DC power in Watts instead of Kilowatts, creating impossible power spikes. These were normalized back to Kilowatts.
3. Inverter Efficiency Correlation Audit: A correlation audit proved that logged inverter efficiency was independent of all physical parameters ($r < 0.02$), behaving as a static constant ($96.99% \pm 0.86%$). We treated efficiency as a fixed system parameter ($97%$) and excluded it from the ML pipeline to prevent overfitting on random sensor noise.
4. Temporal Split: To prevent data leakage, we sorted the dataset chronologically and split it: the first $80%$ of the timeline (spring) for training ($14,332$ rows) and the remaining $20%$ (early summer) for testing ($3,583$ rows).
5. Solar Geometry Features: Added solar zenith angles (computed via pvlib) and direct/diffuse irradiance splits to allow the model to understand the sun's trajectory relative to the tilted panels.
6. Inverter One-Hot Encoding: Handled a systematic $24.8%$ output discrepancy between the 8 physical string inverters across the solar field by one-hot encoding the inverter IDs.

Comparative Model Performance (On Unseen Test Set):

Both machine learning models were evaluated on the chronologically unseen test set:

Model Architecture	R-squared ($R^2$)	MAE (kW)	RMSE (kW)
Random Forest	$0.6802$	$15.0533\text{ kW}$	$21.0272\text{ kW}$
XGBoost (Our Model)	$0.6903$	$14.6817\text{ kW}$	$20.6905\text{ kW}$

---

4. Transposition to Bouskoura, Morocco

Using the NASA POWER API, we retrieved a complete, hourly weather dataset for Bouskoura for the year 2025 ($8,760$ hours).

• Cleaning Sentinel Values: We audited the raw NASA dataset, located 24 corrupted hourly irradiance readings logged as the sentinel -999.0, and resolved them using linear time-series interpolation.
• Zenith Calculations: We used pvlib to calculate the solar zenith angle for every hour of 2025 based on the campus coordinates.
• Daytime Prediction Constraint: To prevent training errors, we filtered the Bouskoura dataset to isolate the $4,015$ active daytime hours (GHI > 50 W/m²). We scaled and predicted only on these daytime hours, and then reconstructed the full 8,760-hour calendar year by mapping the predictions back into a timeline of zeros (representing nighttime).
• System Sizing Scaling: Since our XGBoost model was trained on a single $170\text{ kW}$ SCADA inverter system, we scaled the predictions to match the proposed $1,082.37\text{ kWp}$ campus array: $$\text{Campus Power (kW)} = \frac{\text{Predicted Power (kW)}}{170.0} \times 1,082.37$$

---

5. Sizing & Generation Findings

The transposed machine learning model predicts a realistic, thermally degraded annual energy output of $1,411,959.07\text{ kWh/year}$ (approx. $1.41\text{ GWh/year}$).

A. Expected Sizing Range (With a $\pm6%$ Margin): $$\text{Expected Annual Generation} = \mathbf{1,327,228\text{ to } 1,496,660\text{ kWh/year}}$$

This corresponds to a realistic system yield of $1,304.51\text{ kWh/kWp/year}$.

B. Campus Demand Coverage:

• Current Demand Coverage: This system covers $91.28%$ of the school's 2023 annual consumption ($1,546,749\text{ kWh}$).
• Future Demand Coverage: This system covers $65.30%$ of the school's projected annual consumption ($2,162,213\text{ kWh}$).

C. Monthly Generation Profile:

Model-Predicted Monthly Generation (2025):
January    :  86,790.26 kWh
February   :  87,618.63 kWh
March      : 111,497.59 kWh
April      : 120,230.19 kWh
May        : 146,647.90 kWh
June       : 149,063.21 kWh
July       : 158,681.13 kWh (Summer Solstice Peak)
August     : 152,824.90 kWh
September  : 130,890.91 kWh
October    : 109,158.59 kWh
November   :  88,469.24 kWh
December   :  70,086.52 kWh (Winter Solstice Trough)

---

6. Symmetrical Reproduction Steps

To execute the data pipeline and reproduce the predictions:

1. Clone this repository:

   git clone https://github.com/ALZ-11/campus-energy-optimization
   cd campus-energy-optimization

2. Install all dependencies:

   pip install -r requirements.txt

3. Place the Raw SCADA Files: Download the raw files from the Mendeley Data Repository and place them as:
   • data/raw/Inverter SCADA.csv
   • data/raw/Meteorological data.csv
4. Run the Notebook: Open and execute all cells in notebooks/campus-energy-optimization.ipynb. The notebook will automatically clean the SCADA data, train the XGBoost model, query the live NASA POWER API for Bouskoura's 2025 weather, and generate the final monthly yield forecasts.