Data engineering
Built a reproducible workflow for aligning hourly grid demand, regional observations, weather, calendar variables, and lag features.
CAISO × ELECTRICITY LOAD FORECASTING
Predicting tomorrow’s demand, one hour at a time.
An end-to-end forecasting project that transforms historical grid demand, weather, calendar context, and weekly patterns into hourly electricity-load predictions for California and four subregions.
Team research projectPresented at Millennium Management Global Investment
BEST FINAL520MW MAE
FORECAST WINDOW168 hours
REGIONS5
TOP MODELST-CALNet
RECRUITER SNAPSHOT
What this project demonstrates
Model benchmarking
Compared simple baselines with Random Forest, XGBoost, LSTM variants, and ST-CALNet using chronological evaluation.
Time-series reasoning
Used full-week windows, no-shuffle data splits, train-only scaling, and lag construction to prevent future-data leakage.
Product communication
Translated technical results into operational questions: peak demand, regional specialization, reliability, and deployment tradeoffs.
THE BUSINESS PROBLEM
Electricity must be available before customers ask for it.
Grid operators need an accurate view of future demand to plan generation, manage reserves, and prepare for peak hours. Forecast too low and reliability is at risk. Forecast too high and resources may be scheduled inefficiently.
Reduce the size of forecasting errors
Better forecasts help operators anticipate morning ramps, afternoon cooling demand, and evening peaks.
Explain why the model made its prediction
Simpler models expose coefficients and feature effects, making them easier to audit and communicate.
When is additional complexity worth it?
The project benchmarks each architecture to verify that gains in accuracy are real—not simply the result of a more complicated model.
INTERACTIVE FORECAST LAB
See how an hourly load forecast is built.
Select a region, model, and forecasting strategy. The interface then walks through the same conceptual pipeline used in the project.
STAGE 01
Collect historical electricity demand
Hourly CAISO and subregion demand becomes the foundation of the forecast.
The pipeline preserves chronological order so the model only learns from information that would have existed at prediction time.24-HOUR PORTFOLIO DEMONSTRATION
California Independent System Operator
The system-wide view combines several climates, customer groups, and regional demand patterns.
Forecast ready
34,377 MW
18,863 MW
Demonstration metric calculated from the representative chart above.
6 PM
6 PM
CNN + LSTM + attention
The chart uses representative portfolio data to demonstrate the product experience. Reported project metrics are labeled separately and come from the team’s completed model evaluation.
FROM RAW DATA TO FORECAST
The model does not begin with a neural network.
Reliable forecasting starts with a disciplined data pipeline. The project aligns each observation to the correct hour, region, and historical context before any model is trained.
Hourly grid demand
CAISO and regional load observations provide the target the models learn to predict.
Weather context
Temperature and environmental variables explain cooling and heating behavior.
Time features
Hour, weekday, season, and lagged demand expose repeated daily and weekly patterns.
Chronological split
Training, validation, and test sets remain ordered to prevent information from the future leaking backward.
MODEL INPUTS
What the forecast can see
Classical models receive engineered tabular features. Sequence models also receive an ordered 168-hour history so they can learn transitions across an entire week.
A KEY FINDING
Forecast locally, then add the regions together.
A single CAISO model must compromise across several climates and customer mixes. Separate regional models can specialize before their forecasts are summed into one system-level prediction.
SCE
PGE
SDGE
VEA
SUMCAISOsystem forecast
DIRECT VS. BOTTOM-UP CAISO FORECASTINGLower MAE is better
Linear Regression6.1% improvement
Random Forest8.8% improvement
XGBoost11.6% improvement
01
Specialization: each regional model learns its own climate and demand shape.
02
Error cancellation: some regional over- and under-predictions offset when summed.
MODEL PROGRESSION
Complexity had to earn its place.
Each model answered a different question—from whether simple coefficients were sufficient to whether attention could identify the most useful moments in a full week of history.
01
Classical baseline
Linear Regression
Assigns one coefficient to each weather, calendar, hour, and lag feature. The prediction is a weighted sum of those inputs.
WHY IT MATTERED
Established a transparent baseline and made it possible to measure whether added complexity produced meaningful value.
REPORTED RESULT613.9 MW pooled MAE
02
Tree ensemble
Random Forest
Averages hundreds of decision trees, allowing the forecast to respond differently across temperatures, hours, and regions.
WHY IT MATTERED
Demonstrated that nonlinear feature interactions materially improve forecasting over a linear baseline.
REPORTED RESULT488.3 MW pooled MAE
03
Boosted trees
XGBoost
Builds trees sequentially so each new tree focuses on correcting the residual errors left by the previous ensemble.
WHY IT MATTERED
Created the strongest classical benchmark and provided a high bar for evaluating deep sequence models.
REPORTED RESULT1,014 MW bottom-up CAISO MAE
04
Deep sequence model
Bidirectional LSTM
Processes a 168-hour window in both directions, learning daily and weekly demand patterns from an ordered sequence.
WHY IT MATTERED
Proved that temporal sequence learning adds signal beyond a carefully tuned tree-based benchmark.
REPORTED RESULT827 MW MAE · 3.08% MAPE
05
CNN + LSTM + attention
ST-CALNet
Combines convolutional feature extraction, recurrent sequence learning, and attention over the most relevant historical moments.
WHY IT MATTERED
Delivered the project’s strongest final result by integrating local patterns, long-range temporal context, and selective attention.
REPORTED RESULT520 MW final MAE
CLASSICAL MODEL BENCHMARK
Pooled performance across five regions
These values belong to the pooled classical evaluation and should not be compared as though every later model used the identical aggregation context.
ModelMAERMSEContext
Linear Regression613.9 MW1,071.6 MWPooled across CAISO and four subregions
Random Forest488.3 MW868.9 MWPooled across CAISO and four subregions
XGBoost483.7 MW870.2 MWBest pooled classical MAE
BEST FINAL MODEL
Inside ST-CALNet
The final architecture combines three complementary ideas: a CNN for local patterns, an LSTM for sequence memory, and attention for deciding which historical moments deserve the most weight.
Historical inputs
Seven days of load, weather, lag, and time features.
Convolution
Detects local ramps, spikes, and short-term hourly patterns.
LSTM memory
Learns how demand evolves across daily and weekly sequences.
Attention
Weights the historical moments most useful for the forecast.
Load forecast
Produces the expected future demand in megawatts.
FINAL ST-CALNET RESULT520 MWmean absolute error
Extracts rapid hourly fluctuations and short local patterns.
Tracks how demand evolves across the seven-day sequence.
Focuses the forecast on the most relevant historical moments.
EVALUATION WITHOUT THE JARGON
Four metrics answer four different questions.
How far off are we, on average?
Mean Absolute Error reports the typical miss in megawatts. It is the most intuitive headline metric for this project.
Are there any especially large misses?
Root Mean Squared Error penalizes large errors more heavily, making it useful when peak-hour mistakes are especially costly.
What percentage of demand did we miss?
Mean Absolute Percentage Error normalizes the error, but it must be interpreted carefully for very small regional loads.
How much demand variation is explained?
R² measures how well the forecast captures the overall rises, falls, and recurring patterns in electricity demand.
HONEST LIMITATIONS
A strong metric is not the end of the analysis.
The presentation treats limitations as part of the engineering result rather than hiding them behind the best number.
No future weather forecast
Historical weather helps explain past demand, but production forecasting should include the weather expected during the target horizon.
Next step: forecast-weather integrationDistribution shift
The Bidirectional LSTM’s validation-to-test gap suggests that late-period demand differed from the data used to fit the model.
Next step: rolling retrainingMissing event features
Holidays, wildfire periods, unusual heat events, and changing EV demand can alter load in ways ordinary calendar features miss.
Next step: event-aware featuresComputational cost
ST-CALNet improves accuracy but requires more training and inference resources than the classical models.
Next step: latency and cost benchmarkingFROM NOTEBOOK TO SYSTEM
How the research could become a live forecasting service.
GridStatus.io
Collect the latest CAISO demand observations on a schedule.
Hourly pipeline
Build the newest 168-hour window and future-known time inputs.
Saved model weights
Load the selected model and generate the next forecast curve.
Hourly scoring
Compare predictions with actual load and watch for drift.
PROJECT TAKEAWAYS
The final result is more than one model score.
Nonlinearity matters. Tree ensembles reduced pooled error substantially compared with the linear baseline.
Geography matters. Regional specialization and bottom-up aggregation improved the CAISO forecast.
Sequence matters. Full-week context allowed deep models to learn recurring demand patterns that static features only approximate.
Data still matters most. Better architecture cannot fully compensate for missing future weather or major external events.
CAISO LOAD FORECASTING