About ColdFlow

Why Hathras

Hathras sits in western Uttar Pradesh's potato belt. In 2023-24 the district harvested about 1,362,000 t from 55,200 ha, and output has grown 2.26% a year since 2018-19.

Almost the whole crop is dug in a few weeks from late February. Most of it has to go into cold storage within about two months, so storage capacity, not just production, decides what farmers earn. When stores fill up, the surplus is sold at harvest-glut prices or trucked out of the district.

What ColdFlow does

• Publish the district's area, production and yield series in one place, with downloads.
• Project the next two harvests with a transparent statistical method.
• Model month-by-month cold storage stock against capacity and flag shortfalls.
• Show which development blocks are likely to be under the most strain.
• Run the same analysis on your own figures, and save plans to your account.

Methodology

Growth and yield

Year-on-year change is the percentage change from the previous season. Annual growth (CAGR) is computed between 2018-19 and 2023-24. Yield is production ÷ area.

Forecast

Area, production and yield are each fitted with an ordinary least-squares straight line against the season index (0 = 2018-19). The 95% prediction interval uses the Student-t distribution with n − 2 degrees of freedom. With only six seasons, the interval is wide, and treating the range as the planning figure is safer than the point estimate. Weather, prices and seed availability are not modelled. A damped exponential-smoothing (Holt) forecast is shown alongside for comparison.

Forecast evaluation

Models are tested with a rolling origin: from each season (training on at least three earlier seasons), each model forecasts the next one and two seasons using only data that existed at the time, and the forecast is compared with what happened. The linear trend and smoothing models are compared with three simple baselines: naive (same as last season), drift (last season plus the average yearly change) and the historical mean. ColdFlow reports MAE (mean absolute error, tonnes), WAPE (total absolute error ÷ total actual harvest), bias (total error ÷ total actual, positive means over-forecast) and skill against the naive baseline. With six seasons this gives only three one-season-ahead tests, so the results are labelled indicative; the headline forecast is not switched between models on that evidence. See the current results.

Cold storage model

Storage demand is the harvest × the share stored (75% by default). That volume enters storage following a monthly intake profile (February–April) and leaves following a release profile (May–November). Each month, intake beyond free capacity is counted as unmet demand. Releases are shares of what was actually loaded, since produce that found no space never entered storage. Utilisation is the stock the season would hold with unlimited space ÷ capacity. Status levels are: Adequate below the alert threshold − 15 points, Watch up to the threshold, Strained from the threshold to 100%, and Deficit above 100%.

Holding costs

Electricity cost = tonne-months held × kWh per tonne-month × tariff. Each month's stock is the average of its opening and closing stock. The default 9 kWh per tonne-month and ₹7.5/kWh are placeholders; enter your own figures in the planner.

Redistribution plan

The planner solves a linear program (SciPy's HiGHS solver) in two stages: first it maximises the tonnes that find storage across the blocks, then, among those plans, it minimises the tonne-kilometres hauled, optionally within a maximum haul distance. It uses only block demand, block capacity and straight-line distance between block headquarters (with an optional road factor), so its results inherit the modelled block split and are labelled as a demo. Freight cost is shown only when you enter your own rate per tonne-km. It is a single-period peak model: it doesn't schedule trucks, price individual stores or account for crop quality.

Your own data

On the Plan your data page, the forecast and storage model run on the seasons and calendar you enter. ColdFlow checks that seasons are consecutive and flags yields outside 10–50 t/ha, which usually means mixed-up units. Results are labelled as based on your data and are only as good as the inputs.

Block map

District demand and capacity are split across the seven development blocks with modelled shares. Treat the block layer as illustrative until block-level production and the cold storage registry are added.