Workforce Demand Forecasting

Overview

The Workforce Demand Forecasting model predicts the number of field crew hours required across skill categories over an 18-month forward horizon. It uses a Prophet + XGBoost ensemble architecture, with Prophet modelling temporal structure (trend, seasonality) and XGBoost modelling exogenous business drivers (capex programs, vegetation seasons, storm events).

Model Architecture

# Ensemble structure
class WorkforceForecastEnsemble:
    """
    Stage 1: Prophet models temporal patterns
    Stage 2: XGBoost models residuals using exogenous features
    Final: Blend Prophet base + XGBoost adjustment
    """

    def __init__(self):
        self.prophet_models = {}     # One per skill category
        self.xgb_models = {}         # One per skill category
        self.blend_weights = {}      # Learned blending weights

    def predict(self, horizon_months: int, exog_features: dict) -> pd.DataFrame:
        prophet_forecast = self.prophet_predict(horizon_months)
        xgb_adjustment = self.xgb_predict(exog_features)
        return prophet_forecast + xgb_adjustment

Model Performance

Metric	Value	Notes
MAE	124 crew-hours/month	Absolute error per skill category
MAPE	4.2%	Percentage error
Forecast horizon	18 months	Beyond 12 months, interval widens significantly
Update frequency	Monthly	Retrains on full history + new actuals

Prophet Components

The Prophet base model captures:

Long-Term Trend

Workforce requirements grow with network expansion (capex programs, DER connections). The trend is modelled as a piecewise linear function with change points at regulatory period boundaries.

Annual Seasonality

• Summer peak (Nov–Feb): +35–80% above base for vegetation clearance during bushfire season
• Autumn taper (Mar–Apr): +15% above base for post-summer remediation
• Winter base (May–Aug): lowest demand, focus on planned asset replacement
• Spring ramp (Sep–Oct): pre-season vegetation clearance preparation

Weekly Seasonality

• Monday: peak of planned work (fresh week)
• Friday: reduced (most crews return to depot)
• Saturday/Sunday: emergency response only (small baseline)

XGBoost Exogenous Features

Feature	Lag	Description
approved_capex_qtr	0	Approved capex for next quarter ($M)
vegetation_program_km	0	Planned vegetation clearance km
scheduled_replacements	0	Count of scheduled asset replacements
storm_season_flag	0	Binary: Oct–Mar bushfire/storm season
public_holiday_density	0	Number of public holidays in month
prior_month_actual	1	Prior month actual crew-hours (lag 1)
prior_year_actual	12	Same month prior year (lag 12)

Skill Category Forecasts

The model produces separate forecasts for 6 skill categories:

Skill Category	Description	Forecast Accuracy (MAPE)
HV Linesperson	High-voltage overhead line work	3.8%
LV Linesperson	Low-voltage and service work	4.1%
Cable Jointer	Underground cable work	5.2%
Protection Engineer	Protection relay testing and design	6.1%
Project Manager	Capital project management	5.8%
Field Inspector	Vegetation and asset inspections	3.4%

Prediction Intervals

The model produces P10/P50/P90 prediction intervals:

# Prediction interval calculation
def get_prediction_intervals(forecast: np.ndarray, residuals_std: np.ndarray) -> dict:
    """
    P10 (lower bound): 10th percentile — used for minimum staffing planning
    P50 (base case): median — primary planning figure
    P90 (upper bound): 90th percentile — contingency staffing buffer
    """
    return {
        "P10": forecast - 1.28 * residuals_std,
        "P50": forecast,
        "P90": forecast + 1.28 * residuals_std
    }

Interval width increases with horizon:

• 1–3 months: ±8% (narrow)
• 4–12 months: ±15% (moderate)
• 13–18 months: ±25% (wide)

Skills Gap Analysis

The skills gap is computed by comparing forecast demand against current workforce capacity:

Gap_month_skill = Forecast_demand_month_skill - (FTE_count_skill × Available_hours_per_FTE)

Positive gap = shortage (need contractors or hire) Negative gap = surplus (potential redeployment or reduced contractor spend)

Bushfire Seasonality

Bushfire season creates the most significant demand spikes. The model explicitly incorporates:

1. Historical fire weather index: days above Fire Danger Rating thresholds per region
2. DNSP vegetation risk scores: aggregate risk score drives inspection and clearance demand
3. Post-fire remediation factor: after significant fire events, network rebuilding demand is elevated for 6–12 months

API Endpoints

# 18-month workforce demand forecast
GET /api/dnsp/workforce/forecast?dnsp=ausgrid&horizon=18m

# Specific skill category
GET /api/dnsp/workforce/forecast?dnsp=ergon&skill=hv_linesperson&horizon=12m

# Skills gap analysis
GET /api/dnsp/workforce/skills-gap?dnsp=sa_power_networks&horizon=12m

# Forecast with prediction intervals
GET /api/dnsp/workforce/forecast?dnsp=energex&include_intervals=true