Data imputation in in situ-measured particle size distributions by means of neural networks

Abstract. In air quality research, often only size-integrated particle mass concentrations as indicators of aerosol particles are considered. However, the do not provide sufficient information to convey full story fractionated size distribution, in which different diameters (Dp) able deposit differently on respiratory system and cause various harm. Aerosol distribution measurements rely a variety techniques classify measure distribution. From raw data ambient is determined utilising suite inversion algorithms. problem quite ill-posed challenging solve. Due instrumental insufficiency limitations, imputation methods for great significance fill missing gaps or negative values. The study at hand involves merged from scanning mobility sizer (NanoSMPS) an optical (OPS) covering distributions 0.01 0.42 µm (electrical equivalent size) 0.3 10 (optical meteorological parameters collected urban background region Amman, Jordan, period 1 August 2016–31 July 2017. We develop evaluate feed-forward neural network (FFNN) approaches estimate number particular bin with (1) parameters, (2) concentration other bins (3) both above input variables. Two layers 10–15 neurons found be optimal option. Worse performance observed lower edge (0.01<Dp<0.02 µm), mid-range (0.15<Dp<0.5 µm) upper (6<Dp<10 µm). For edges ends, neighbouring limited, detection efficiency by corresponding instruments compared bins. A distinct drop over overlapping due deficiency merging algorithm. Another plausible reason poorer finer that they more effectively removed atmosphere coarser so relationships between variables small dynamic. An observable overestimation also early morning ultrafine followed underestimation before midday. winter, possible sensor drift interference artefacts, estimation good seasons. FFNN approach using 5 min (R2= 0.22–0.58) shows results than longer time resolution 0.66–0.77). can serve alternative way replace numbers dataset thanks its high accuracy reliability 0.97–1). This negative-number filling maintain symmetric errors complement existing built-in algorithm instruments.