Documentation

What it shows: Stacked area chart of each generation resource's available capacity (24 hours) overlaid with a demand line.

How to read it:

• The colored stacked areas represent each resource type's contribution (solar, wind, gas, hydro, nuclear, etc.)
• The white dashed line shows actual demand
• The gap between the top of the stack and the demand line is the effective headroom (spare capacity)
• Use the date slider at the top to explore any day from 2020–2025

Key insight: On sunny spring days with low demand, solar pushes total capacity far above demand — this is where overgeneration and negative prices occur.

What it shows: For each hour of the day, the distribution of effective headroom across all days in the dataset.

Bands:

• p5 & p95: Outer bounds — 90% of days fall within these lines
• p25 & p75: Inner shaded band — the middle 50% of days
• p50: The median headroom

Key insight: Headroom is tightest during evening peak hours (5–8 PM) when solar drops off but demand stays high. The p5 line dipping toward or below zero means 5% of days have dangerously low spare capacity at that hour.

What it shows: Each dot is one hour from the dataset. X-axis = net load (demand minus all non-gas generation), Y-axis = LMP (wholesale electricity price).

Color: Time of day (yellow = daytime solar hours, blue/purple = night hours)

Dot size: Scaled by CA natural gas citygate price — larger dots = more expensive gas months

Key insight: When net load is negative (surplus), LMP drops to zero or negative — the grid is paying to offload excess generation. When net load is high, LMP rises sharply as expensive peaker plants come online. Gas price shifts the entire curve vertically.

What it shows: The baseline headroom percentile bands (shaded, from the chart above) with additional dashed lines showing the median headroom after adding each data center profile.

Profile lines:

• Blue dashed: Flat DC — uniform headroom reduction at all hours
• Green dashed: Camel DC — more impact during solar hours, less during evening peak
• Red dashed: Night-Heavy DC — worst impact during evening peak when headroom is already tightest
• Purple dashed: Custom DC (when active)

Key insight: The camel profile's line stays closest to the baseline during evening peak hours (4–9 PM) because it consumes less at those times, preserving reliability when it matters most.

What it shows: Color-coded stat cards comparing Baseline, Flat, Camel, Night-Heavy, and Custom profiles across three metrics:

Arrows: Green up-arrows indicate improvement vs baseline; red down-arrows indicate degradation.

Key insight: The camel profile increases utilization (good) while adding fewer stress hours than the flat or night-heavy profiles.

Methodology: Empirical LMP-Headroom Model

Unlike a simple weighted average that assumes prices don't change when load is added, this analysis uses an empirical relationship learned from 6 years of CAISO data to model how adding DC load shifts grid headroom and thereby affects LMP.

Why Not Use Historical LMP Directly?

A naive approach would compute: cost = Σ(dc_load[h] × historical_LMP[h]) / Σ(dc_load[h]). This produces misleading results because:

• Adding 500-2000 MW of load reduces headroom (available capacity minus demand minus reserves)
• Lower headroom → generators move up the supply curve → higher wholesale prices
• Using historical LMP assumes the DC doesn't affect prices (only valid for very small loads)

Result: All DC profiles would show nearly identical costs (~$35/MWh), with no price signal for flexibility value.

Building the LMP-Headroom Lookup Table

We extracted the empirical relationship between effective headroom and LMP from ~52,080 hourly observations (2020-2025):

Step 1: Define 6 headroom bins (based on scatter plot analysis)

Step 2: Normalize LMP by gas price

Removes gas price volatility (2020-2021: ~$2-3/MCF vs 2022: ~$7-9/MCF) to extract the structural relationship between headroom and pricing.

Step 3: Build 24×6 lookup matrix

For each hour-of-day (0-23) and headroom bin (0-5):

1. Collect all historical observations that fall into that (hour, bin) combination
2. Filter outliers (normalized LMP > 200 MCF^-1)
3. Compute median normalized LMP for the bin

This captures:

• Nonlinear response: As headroom → 0, prices spike exponentially
• Hour-of-day effects: Evening hours (8pm) have higher prices than midday at same headroom
• Stable pricing zones: 20-40 GW headroom shows consistently low prices

Applying the Model

For each day with price data and each hour:

Baseline (No DC):

With DC Profile:

Note: Reserves are recomputed with DC load since they scale with total demand.

Chart Display

Bar chart: Single bar per profile showing adjusted LMP ($/MWh). No component breakdown (MEC/MCC/Loss/GHG) because the lookup table provides total LMP only.

Baseline bar: Simple average LMP under historical grid conditions (no DC load). Provides reference for what grid prices were before any DC is added.

Stat cards below: Show per-profile $/MWh, annual cost ($M/year), and savings vs Flat (not Baseline). Color borders match profile colors.

Key Results

How it works:

1. For each of 2,192 days, compute the minimum hourly headroom (baseline)
2. Sort all days by minimum headroom (ascending — worst days first)
3. Take the 100 most-stressed days
4. Display a heatmap: rows = days, columns = 24 hours, color = headroom

Color scale:

• Red = negative or very low headroom (<0 GW)
• Yellow = moderate headroom (~10 GW)
• Green = adequate headroom (~15–20 GW)
• Blue = ample headroom (>20 GW)

Tabs: Switch between Baseline, +Flat DC, +Camel DC, +Night DC, and +Custom DC to see how each profile changes the heatmap pattern. Hover over any cell for exact headroom values.

Summary cards: Show average minimum headroom, worst single hour, and stress hour count for the selected profile — all computed on the 100 stress days only.

What is surplus? Hours when all non-gas generation exceeds demand:

Where non_gas_gen = Solar + Wind + Hydro + Nuclear + Geothermal + Biomass + Other Thermal.

When surplus > 0, even with ALL gas plants shut off, the grid overproduces. This energy is typically curtailed (mostly solar).

Chart: Monthly bar chart showing surplus (gray line) and how much each DC profile absorbs (colored bars). Values are annualized (divided by years of data).

Absorption formula:

Camel DCs absorb the most because they peak during solar hours — exactly when surplus is highest.

Why include hydro in non-gas generation?

• During spring snowmelt, environmental flow minimums make hydro largely inflexible
• CAISO's worst overgeneration events coincide with high hydro + high solar
• Without hydro, the model produces near-zero surplus, contradicting real CAISO data
• Including hydro produces ~3.1 TWh/year surplus, close to CAISO's actual ~3.4 TWh curtailed in 2025

Methodology: Dispatch-Stack Proxy

For each hour, we determine the marginal generator by walking the gas merit order:

1. Compute gas residual: gasResidual = max(0, demand - nonGasGen)
2. Walk merit order (cheapest first):
   • Gas CCGT capacity: if gasResidual fits → marginal rate = 0.41 tCO²/MWh
   • Gas Steam → 0.55 tCO²/MWh
   • Gas ICE → 0.50 tCO²/MWh
   • Gas Peaker → 0.63 tCO²/MWh
3. If gasResidual = 0 (surplus): marginal rate = 0 tCO²/MWh — adding load avoids RE curtailment rather than causing gas combustion

Rates derived from standard heat rates × EPA emission factor for natural gas (53.06 kg CO²/MMBtu).

Chart (dual axis):

• Left axis (solid lines): Average hourly CO² emissions (tCO²/hr) for each DC profile — shows WHERE each profile's emissions come from across the day
• Right axis (dashed gray): Grid marginal emission rate (tCO²/MWh) — the "emissions duck curve" showing that midday solar hours have near-zero marginal emissions

Stat cards: Show per-profile weighted average emission rate (tCO²/MWh), annual CO² (ktCO²/yr), and savings vs Flat.

Purpose: While preset profiles (Flat, Camel, Night-Heavy) demonstrate general patterns, this calculator finds the theoretically best load allocation across all 24 hours for each day, given real grid conditions and empirical LMP-headroom relationships.

Methodology: Iterative Greedy Optimization with Constraints

For a data center of size avgMw, the optimizer allocates avgMw × 24 MWh/day of energy across 24 hours to minimize cost:

1. Initialization: Set optLoad[h] = 0 for all hours (h = 0..23)
2. Increment sizing:
   • avgMw ≤ 1,000 MW → 50 MW increments
   • avgMw ≤ 2,500 MW → 100 MW increments
   • avgMw > 2,500 MW → 200 MW increments

   Larger increments for bigger DCs reduce computation time while maintaining accuracy.

3. Iterative allocation: Repeat until sum(optLoad) ≈ avgMw × 24:
   1. For each hour h:
      • Skip if optLoad[h] + INCREMENT_MW > 3 × avgMw (hourly capacity constraint)
      • Compute test demand: testDemand[h] = demand[h] + optLoad[h] + INCREMENT_MW
      • Compute test reserves: testReserves = max(0.06 × max(testDemand), 3500)
      • Compute headroom: headroom = available[h] - testDemand[h] - testReserves[h]
      • Lookup marginal LMP from empirical 24×6 lookup table: LMP[h][bin(headroom)]
      • Compute marginal cost: cost = INCREMENT_MW × LMP
   2. Find hour h* with lowest marginal cost
   3. Allocate: optLoad[h*] += INCREMENT_MW
4. Rescaling: After allocation, rescale to exactly match target:
   scale = (avgMw × 24) / sum(optLoad) optLoad[h] = optLoad[h] × scale for all h
5. Cost computation: For each day's optimized profile, recompute actual cost using the adjusted headroom-based LMP (same method as other profiles)

Why Greedy Works

The greedy approach (always choosing the lowest-cost hour for the next increment) is near-optimal because:

• The marginal LMP curve is convex in headroom — adding load to tight hours raises LMP more than adding to loose hours
• By repeatedly choosing the cheapest hour, the optimizer naturally spreads load to multiple low-cost hours rather than overloading a single hour
• The iterative approach accounts for feedback effects: as we add load to an hour, its headroom decreases and LMP rises, making other hours more attractive

Results Display

Chart: Dual-bar chart showing:

• Gray bars: Average optimized profile across all 2,170+ days (reference)
• Purple bars: Optimized profile for the currently selected date (synced with date scrubber)

The chart's y-axis is fixed across all dates (set to 1.2× the maximum hourly load across all optimized days) to make day-to-day variation visible.

Scorecard: Four cost comparison cards:

• Baseline (Flat 500 MW): Flat profile at 500 MW (reference grid condition from main analyses)
• Flat: Flat profile at the selected DC size
• Camel (Solar-Following): Preset camel profile at selected DC size and flex ratio
• ⚡ Optimized: The theoretically best allocation

Each card shows:

• Average cost ($/MWh)
• Annual cost ($M/yr at selected DC size)
• Cost difference vs Optimized (should be ≥ 0 for all other profiles, since Optimized is minimum)

Key Insights

Performance

Typical computation times (client-side JavaScript):

Computation is opt-in (triggered by "Calculate" button) to avoid slowing down the main dashboard.