Before any math or machine learning, you need to understand the two real-world problems the papers tackle. This module assumes zero background in energy systems or robotics — it is written for a smart newcomer.
A utility-scale photovoltaic (PV) plant is a field of solar panels feeding electricity into the national grid. The one in Paper 1 is huge: rated at 75 megawatts (MW), with about 312,500 individual PV modules covering roughly a square kilometre in South Africa's sunny Northern Cape. The panels' DC power flows into 84 inverters — devices that convert DC to grid-ready AC — and then to the grid connection point.
The grid must balance supply and demand every second. But PV output is intermittent: a passing cloud can slash a plant's output within minutes. Grid operators therefore distrust solar unless they can see it coming. That is power forecasting: predicting the plant's output ahead of time so operators can plan reserves, balance the grid, and trade energy. Paper 1 targets intra-day forecasts, 1 to 6 hours ahead, at 15-minute resolution — 21 future values predicted at once — using only on-site historical measurements (no external weather forecasts).
Nearly all published forecasting models are trained on the plant's total output — a macro-level model that treats the whole farm as one big panel. But a large plant is not uniform: wind cools some rows more than others, clouds and terrain shade parts of the field, so different segments behave differently — up to 3% output variation between inverters just from wind direction. Paper 1 asks: would 84 small "low-level" forecasts (one per inverter), added together, beat one big macro forecast? And: can standard deep learning models capture that low-level behaviour at all?
Before a robot arm is built, engineers create a CAD model (a precise 3D blueprint) and derive its kinematics — the equations linking each motor's joint angle to where the arm's parts end up in space. With those equations a robot can plan: "to put my fingertip there, set my joints to these angles." This works, but it is hand-built, rigid, and wrong the moment the robot is damaged, worn, or modified.
A child learns their body not from blueprints but by watching themselves move — in a mirror, for instance. Paper 2 gives a robot the same ability: the robot waves its arm around randomly (motor babbling) in front of a single ordinary 2D camera, and from those video frames alone — no 3D scanners, no measuring tape, no kinematic equations — it learns a self-model: an internal, queryable simulation of its own shape and movement.
The learned model — the free-form kinematic self-model (FFKSM) — answers a simple query: "Given my current joint angles, is the 3D point (x, y, z) occupied by my body, and would the camera see it?" Ask that question for thousands of points and you can render the robot's predicted silhouette, plan collision-free motions, and control the arm — all through the model. It is inspired by NeRF, a computer-vision technique for learning 3D scenes from 2D images (Module 6). Training compares the model's rendered silhouette to the actual camera image; nothing else is needed.
A solar farm and a robot arm look unrelated, but the exam machinery underneath is largely shared:
| Shared concept | Paper 1 (Solar) | Paper 2 (Robot) |
|---|---|---|
| Core task | Regression: predict future power values | Regression: predict occupancy/visibility of 3D points |
| Input data | Sensor time series (power, irradiance, weather) | Video frames + joint angles |
| Neural architectures | FFNN, LSTM, GRU | Coordinate/kinematic encoders + predictive module (fully connected, ReLU) |
| Loss function | MSE on power output | MSE on rendered vs real silhouette pixels |
| Optimizer | Adam, lr 1×10⁻⁴ | Adam (also used for control, lr 0.04) |
| Data headache | Missing/faulty sensor readings | Class imbalance: mostly-black images |
| Honest evaluation | Bootstrap confidence intervals, weather-type breakdown | RS/NN baselines, 2,000-sample test set |
| Adaptation story | Clustering to avoid retraining 84 models | Fine-tuning to recover from damage |
Master the shared machinery once (Modules 1–8), then each paper becomes a case study (Modules 9–10) rather than a mountain.